Sperber Wilson Relevance 1996 PDF

Relevance
Communication and Cognition

Second Edition
DAN SPERBER
and
DEIRDRE WILS0N'
I]
BLACI<WELL
Oxford UK &- Cambridge USA
Copyright Dan Sperber and Deirdre Wilson, 1986, 1995

The right ofDan Sperber and Deirdre Wilson to be identified as authors of this work has been
asserted in accordance with the Copyright, Designs and Patents Act 1988.
First published 1986
Reprinted 1986, 1988, 1990, 1993
Second edition 1995
Reprinted 1996 (twice)
Blackwell Publishers Ltd
108 Cowley Road
Oxford OX4 IJF, UK
Blackwell Publishers Inc.
238 Main Street
Cambridge, MassacllUSetts 02142, USA
All rights reserved. Except for the quotation of short passages for the purposes
of criticism and review, no part of this publication may be reproduced, stored
in a retrieval system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without the prior permission
of the publisher.
Except in the United States of America, this book is sold subject to the condition
that it shall not, by way of trade or otherwise, be lent, re-sold, hired out, or
otherwise circulated without the publisher's prior consent in any form of binding
or cover other than that in which it is published and without a similar condition
including this condition being imposed on the subsequent purchaser.
British Library Cataloguing in Publication Data

A CIP catalogue record for this book is available from the British Library
Library o/Congress Cataloging in Publication Data

Sperber, Dan.
Relevance: communication and cognitionJDan Sperber and Deirdre Wilson.
p. cm.
Includes bibliographical references and index.
ISBN 0-631-19878-4 (pbk: alkpaper)
1. Oral communication-Psychological aspects. 2. Relevance.
3. Inference. 4. Cognition. 1. Wilson, Deidre. IT. Title.
BF637.C45S655
1995
95-21627
302.2-dc20
CIP
Printed and bound in Great Britain by T. J. Press Ltd, Padstow, Cornwall

This book is printed on acid-free paper
Contents
Preface to Second Edition

List of symbols
1 Communication
The code model and the semiotic approach to communication
Decoding and inference in verbal comprehension
The mutual-knowledge hypothesis
Grice's approach to 'meaning' and communication
Should the code model and the inferential model be amalgamated?
Problems of definition
Problems of explanation: Grice's theory of conversation
Cognitive environments and mutual manifestness
Relevance and ostension
Ostensive-inferential communication
11 The informative intention
12 The communicative intention
1
2
3
4
5
6
7
8
9
10
2 Inference
1
2
3
4
5
6
7
Non -demonstrative inference

Logical forms, propositional attitudes and factual assumptions
Strength of assumptions
Deductive rules and concepts
The deductive device
Some types of deduction
Contextual effects: the role of deduction in non-demonstrative
inference
3 Relevance
1 Conditions for relevance
2 Degrees of relevance: effect and effort
3 Is the context given or chosen?
vu
IX
1
3
9
15
21
24
28
31
38
46
50
54
60
65
65
71
75
83
93
103
108
118
118
123
132
Contents
VI
4
5
6
7
8
A choice of contexts
Relevance to an individual
The relevance of phenomena and stimuli
The principle of relevance
How relevance theory explains ostensive-inferential
communication
4 Aspects of verbal communication

1 Language and communication
2 Verbal communication, explicatures and implicatures
3
4
5
6
7
8
9
10
The identification of propositional form &

The identification of implicatures
Propositional form and style: presuppositional effects
Implicatures and style: poetic effects
Descriptive and interpretive dimensions of language use
Literalness and metaphor
Echoic utterances and irony
Speech acts
137
142
151
155
163
172
172
176
183
193
202
217
224
231
237
243
Postface
255
Notes to First Edition
281
Notes to Second Edition
293
Notes to Postface
295
Bibliography
299
Index
321
In this book, first published nine years ago, we present a new approach
to the study of human communication. This approach (outlined in
chapter 1) is grounded in a general view of human cognition (developed
in chapters 2 and 3). Human cognitive processes, we argue, are geared
to achieving the greatest possible cognitive effect for the smallest possible
processing effort. To achieve this, individuals must focus their attention
on what seems to them to be the most relevant information available.
To communicate is to claim an individual's attention: hence to communicate is to imply that the information communicated is relevant. This
fundamental idea (developed in chapter 3), that communicated information comes with a guarantee of relevance, is what in the First Edition
we called the principle of relevance and what we would now call the
Second, or Communicative Principle of Relevance (see the Postface to
this Second Edition). We argue that this principle of relevance is essential
to explaining human communication, and show (in chapter 4) how it is
enough on its own to account for the interaction of linguistic meaning
and contextual factors in utterance interpretation.
Here is how this book came about. In 1975, Deirdre Wilson published
Presuppositions and Non-Truth-Conditional Semantics and Dan Sperber
published 'Rudiments de rhetorique cognitive', a sequel to his Rethinking Symbolism. In these works, we were both turning to pragmatics the study of contextual factors in verbal communication - but from
different perspectives: Deirdre Wilson was showing how a number of
apparently semantic problems could be better solved at a pragmatic
level; Dan Sperber was arguing for a view of figures of speech rooted in
pragmatics. We then formed the project of writing, in a few months, a
joint essay which would cover, at least programmatically, the ground
between our two vantage points and show the continuities and discontinUlt1es between semantics, pragmatics and rhetoric. Work did not
proceed according to plan. We got involved in carrying out the
VUl
programme we had merely intended to outline. The months became

years. The projected essay became a series of papers and the present
book.
This Second Edition preserves the text of the original, except for the
correction of typographical errors, removal of obvious mistakes and
inconsistencies, updating of existing references, and addition of a few
explanatory notes. In a new Postface, we sketch the main developments
in the theory since the First Edition was published, and argue for some
revisions both of formulation and of substance.
A number of people, who have helped us by their encouragements
and criticisms, bear part of the responsibility for our failure to stick to
our original plan of writing a short, programmatic sketch: Scott Atran,
Regina Blass, Michael Brody, Sylvain Bromberger, Annabel Cormack,
Martin Davies, Sue George, Paul Grice, Ernst-August Gutt, Sam
Guttenplan, Jill House, Pierre J acob, PhilJ ohnson-Laird, Aravind J oshi,
Jerry Katz, Stephen Levinson, Rose MacLaran, George A. Miller, Dinah
Murray, Stephen Neale, Yuji Nishiyama, Ellen Prince, Anne Reboul,
Fran<;ois Recanati, Michael Rochemont, Nicolas Ruwet, Dorota
Rychlik, Tzvetan Todorov, Charles Travis and Bonnie Webber. Dan
Sperber is particularly grateful to Monique Canto-Sperber, Catherine
Cullen and Jenka and Manes Sperber; and Deirdre Wilson to her
colleagues Diane Blakemore, Robyn Carston, Ruth Kempson and Neil
Smith, and especially her husband, Theodore Zeldin. For this Second
Edition, we have benefited from the comments, suggestions and support
of the members of the Relevance e-mail group.
List of symbols
P, Q
U
A
C
I
P
1,2
a, b
individual assumptions
an utterance
a set of assumptions made manifest by an utterance
a set of contextual assumptions
a set of assumptions the communicator intended to make manifest
a set of newly presented assumptions
notes to First Edition
notes to Second Edition
1
Communication
How do human beings communicate with one another? For verbal

communication at least, there is a sort of folk answer, suggested by a
variety of metaphors in everyday use: 'putting one's thoughts into words',
'getting one's ideas across', 'putting one's thoughts down on paper', and
so on.! These make it sound as if verbal communication were a matter of
packing a content (yet another metaphor) into words and sending it off, to
be unpacked by the recipient at the other end. The power of these figures
of speech is such that one tends to forget that the answer they suggest
cannot be true. In writing this book, we have not literally put our thoughts
down on paper. What we have put down on paper are little dark marks, a
copy of which you are now looking at. As for our thoughts, they remain
where they always were, inside our brains.
Suppose it were physically possible to transport thoughts from one
brain to another, as programs and data stored on a magnetic disk can be
transported from one computer to another: then communication would
be unnecessary (whether it might still be useful, for reasons of speed or
economy, is another matter). But thoughts do not travel, and the effects of
human communication cannot be achieved by any other means.
Communication is a process involving two information-processing
devices. One device modifies the physical environment of the other. As a
result, the second device constructs representations similar to representations already stored in the first device. Oral communication, for instance,
is a modification by the speaker of the hearer's acoustic environment, as a
result of which the hearer entertains thoughts similar to the speaker's own.
The study of communication raises two major questions: first, what is
communicated, and second, how is communication achieved?
What is communicated? Meanings, information, propositions,
thoughts, ideas, beliefs, attitudes, emotions, are some of the answers
which have been proposed. More than one of them may well be true.
Communication
Certainly, what is communicated by a religious ritual is quite different

from what is communicated by a list of stock-exchange rates. Even within
the domain of verbal communication, a poem and a legal document seem
to communicate profoundly different things. Nonetheless, we will argue
in section 11 of this chapter that there is a general answer to this question.
For the time being, we will talk quite informally of the communication
of thoughts, assumptions, or information. By thoughts, we mean conceptual representations (as opposed to sensory representations or emotional
states). By assumptions, we mean thoughts treated by the individual as
representations of the actual world (as opposed to fictions, desires, or
representations of representations). Some authors (e.g. Dretske 1981) use
the terms 'information' and 'inform' to talk only of the representation and
transmission of facts; for them, all information is by definition true. We
will use the terms more broadly, treating as information not only facts, but
also dubious and false assumptions presented as factual. In section 8, we
will characterise information more precisely. In chapter 2, we will
consider the structure of thoughts and assumptions in some detail.
Even more important than the question of what is communicated is the
question of how communication is achieved. How can a physical stimulus
bring about the required similarity of thoughts, when there is no similarity
whatsoever between the stimulus and the thoughts it brings into
correspondence? Here again, it is worth considering whether there is a
single, general answer. Should there be - can there be - a general theory of
communication? Most authors, insofar as they are aware of the issue at all,
seem to think that there can, and should.
Let us approach this question in terms of another. Clearly, no one
would waste much time trying to invent a general theory of locomotion.
Walking should be accounted for in terms of a physiological model, plane
flight in terms of an engineering model. While it is true that both walking
and plane flight fall under the same physical laws, these laws are much too
general to constitute a theory of locomotion either. Thus, locomotion is
either too general or not general enough to be the object of an integrated
theory. It is worth considering whether this might not be the case for
communIcation too.
There seems to be a general agreement that there can, and should, be a
general theory of communication. From Aristotle through to modern
semiotics, all theories of communication were based on a single model,
which we will call the code model. According to the code model,
communication is achieved by encoding and decoding messages. Recently, several philosophers, notably Paul Grice and David Lewis, have
proposed a quite different model, which we will call the inferential model.
According to the inferential model, communication is achieved by
producing and interpreting evidence.
The code model and the semiotic approach
The code model and the inferential model are not incompatible; they
can be combined in various ways. The work of pragmatists, philosophers
of language and psycholinguists over the past twenty years has shown that
verbal communication involves both coding and inferential processes.
Thus both the code model and the inferential model can contribute to the
study of verbal communication. However, it is usually assumed that one
of the two models must provide the right overall framework for the study
of communication in general. Most authors take for granted that a proper
theory of communication should be based on the familiar code model; a
few philosophers seem tempted to develop the inferential model into an
inferential theory of communication.
Against these reductionist views, we maintain that communication can
be achieved in ways which are as different from one another as walking is
from plane flight. In particular, communication can be achieved by coding
and decoding messages, and it can be achieved by providing evidence for
an intended inference. The code model and the inferential model are each
adequate to a different mode of communication; hence upgrading either to
the status of a general theory of communication is a mistake. Both coded
communication and inferential communication are subject to general
constraints which apply to all forms of information processing, but these
are too general to constitute a theory of communication either.
Some modes of locomotion involve the interaction of quite different
mechanisms: bicycle riding, for instance, involves both physiology and
engineering. Similarly, verbal communication involves both code and
inferential mechanisms. In trying to construct an adequate description of
these two types of mechanism and their interaction, it is important to
realise that they are intrinsically independent of one another, and that
communication in general is independent of either.
In sections 1 to 3 of this chapter we will discuss the code theory, and in
sections 4 to 7, the inferential theory. In discussing the views of code and
inferential theorists, our aim is to contrast two extreme approaches so as to
map out the full range of available choices; it is not to do justice to those
who have defended subtly qualified, or cautiously vague, versions of
either. In sections 8 to 12 of this chapter and in chapters 2 and 3, we will
propose what we hope is an improved inferential model. However, we do
not regard this model as the basis for a general theory of communication.
In chapter 4, we will show instead how it can be combined with a code
model to provide an explanatory account of verbal communication.
1 The code model and the semiotic approach to communication
A code, as we will use the term, is a system which pairs messages with
signals, enabling two information-processing devices (organisms or
Communication
machines) to communicate. A message is a representation internal to the

communicating devices. A signal is a modification of the external
environment which can be produced by one device and recognised by the
other. A simple code, such as the Morse code, may consist of a
straightforward list of message-signal pairs. A more complex code, such
as English, may consist of a system of symbols and rules generating such
paIrs.
A widely quoted2 diagram of Shannon and Weaver (1949), slightly
adapted in figure 1, shows how communication can be achieved by use of a
code:
received
received
signal
I
I
nOIse
Figure 1
This diagram shows how a message originating in an information source
can be duplicated at a destination as the result of a communication process.
For instance, the source and the destination could be telecommunications
employees, the encoder and the decoder telex machines, the channel an
electric wire, the message a text, i.e. a series of letters, and the signal a series
of electrical impulses. The message is typed by the source on the encoder's
keyboard. The encoder contains a code which associates each letter to a
distinct pattern of electrical impulses. The encoder sends these impulses
through the channel to the decoder. The decoder contains a duplicate of
the encoder's code, and uses it to deliver to the destination the series of
letters and signs associated by the code to the electrical impulses it has
received.
Communication is achieved by encoding a message, which cannot
travel, into a signal, which can, and by decoding this signal at the receiving
end. Noise along the channel (electrical disturbances in our example) can
destroy or distort the signal. Otherwise, as long as the devices are in order
and the codes are identical at both ends, successful communication is
guaranteed.
In this example, the communicating devices are neither the telecommunications employees nor the telex machines but the man-machine pairs
5
on both sides. This apparent complication is, in fact, illuminating. It
shows what the relevant internal structure of any device capable of coded
communication would have to be. Consider the case of honey bees. Von
Frisch (1967) has shown that bees can encode into flight patterns (their
'dance') what they have learnt about the location of nectar, so that other
bees can decode the information and find the nectar in their turn. To
account for this communicative ability, bees must be seen as containing
two information-processing sub-devices: a memory (which constitutes
the 'source' on the one side and the 'destination' on the other) in which
plans for flying to~ards a. supp.ly of nectar can. b.e store~, and an
encoder-decoder devIce whlCh paIrs messages conslstmg of flIght plans
with signals consisting of dances.
It may seem that a similar model could be proposed for human. verbal
communication, as shown in figure 2:
thought
~---.
I ,------,
central
I
thought
linguistic
encoder
processes
received
acoustic
signal
acoustic
signal
f--f---+I
air
SPEAKER
I
I
I
received
thought
I
~--~I~--~
I central
thought
processes
HEARER
nOIse
Figure 2
Here the source and the destination are central thought processes, the
encoder and the decoder are linguistic abilities, the message is a thought,
and the channel is air which carries an acoustic signal. There are two
assumptions underlying this proposal: the first is that human languages,
such as Swahili or English, are codes; the second is that these codes
associate thoughts to sounds.
While Shannon and Weaver's diagram is inspired by telecommunications technology, the basic idea is quite old, and was originally proposed
as an account of verbal communication. To give just two examples:
Aristotle claimed that 'spoken sounds are symbols of affections in the
soul', which are themselves 'likenesses of actual things' (Aristotle, De
Interpretatione: 43). In our terms, he claimed that utterances encode
assumptions. Arnauld and Lancelot in their famous Grammaire de
Port-Royal describe language as
Communication
the marvellous invention of composing out of 25 or 30 sounds that
infinite variety of words, which tho' they have no natural resemblance to the operations of the mind, are yet the means of unfolding
all its secrets, and of disclosing unto those, who cannot see into our
hearts, the variety of our thoughts, and our sentiments upon all
manner of subjects.
Words therefore may be defined, distinct and articulate sounds,
made use of by men as signs, to express their thoughts. (Arnauld and
Lancelot, Grammaire de Port-Royal: 22)
The view of linguistic communication as achieved by encoding

thoughts in sounds is so entrenched in Western culture that it has become
hard to see it as a hypothesis rather than a fact. Yet the code model of
verbal communication is only a hypothesis, with well-known merits and
rather less well-known defects. Its main merit is that it is explanatory:
utterances do succeed in communicating thoughts, and the hypothesis
that they encode thoughts might explain how this is done. Its main defect,
as we will shortly argue, is that it is descriptively inadequate: comprehension involves more than the decoding of a linguistic signal.
The semiotic approach to communication (as Peirce called it and we will
call it ourselves), or the semiological approach (as Saussure and his
followers called it), is a generalisation of the code model of verbal
communication to all forms of communication. Todorov (1977) dates it
back to Augustine, who approached the study of grammar, logic, rhetoric
and hermeneutics within the unifying framework of a theory of signs.
Systems of signs were seen as governing not just the ordinary verbal
communication of thoughts, but also the poetic effects of tropes,
communication by gestures, religious symbols and rites, and the interpretation of sacred texts.
From a semiotic point of view, the existence of an underlying code is the
only possible explanation of how communication is achieved. Here is how
the psychologist Vygotsky formulated this 'axiom':
That understanding between minds is impossible without some
mediating expression is an axiom for scientific psychology. In the
absence of a system of signs, linguistic or other, only the most
primitive and limited type of communication is possible. Communication by means of expressive movements, observed mainly
among animals, is not so much communication as spread of affect ...
Rational, intentional conveying of experience and thought to others
requires a mediating system, the prototype of which is human
speech. (Vygotsky 1962: 6)
Whenever communication is observed, an underlying system of signs is

postulated, and the task of the semiotician is seen as that of reconstructing
it. Saussure's formulation of the programme is well known:
Language is a system of signs that express ideas, and is therefore
comparable to a system of writing, the alphabet of deaf-mutes,
symbolic rites, polite formulas, military signals, etc. But it is the
most important of all these systems.
A science that studies the life of signs within society is conceivable ...
I shall call it semiology. (Saussure 1974: 16)
The semiotic programme has been enthusiastically adopted by a
number of linguists, literary theorists, psychologists, sociologists and
anthropologists. Here is an anthropologist's endorsement:
I shall assume that all the various non-verbal dimensions of culture,
such as style in clothing, village lay-out, architecture, furniture,
food, cooking, music, physical gestures, postural attitudes and so on
are organised in patterned sets so as to incorporate coded information in a manner analogous to the sounds and words and sentences of
a natural language. I assume therefore it is just as meaningful to talk
about the grammatical rules which govern the wearing of clothes as it
is to talk about the grammatical rules which govern speech
utterances. (Leach 1976: 10)
The recent history of semiotics has been one of simultaneous institutional success and intellectual bankruptcy. On the one hand, there are now
departments, institutes, associations, congresses and journals of semiotics.
On the other, semiotics has failed to live up to its promises; indeed, its
foundations have been severely undermined. This is not to deny that many
semioticians have done invaluable empirical work. However, it does not
follow that the semiotic framework has been productive, let alone
theoretically sound; merely that it has not been entirely sterilising, or that
it has not been strictly adhered to in practice. 3
Saussure expected that 'the laws discovered by semiology will be
applicable to linguistics, and the latter will circumscribe a well-defined
area within the mass of anthropological facts' (1974: 16). What actually
happened was that for the few decades in which structuralist linguistics
flourished, the semiotic program was taken seriously and spelled out in
more detail. Linguists such as Hjelmslev (1928, 1959) and Kenneth Pike
(1967) developed ambitious terminological schemes as tools for carrying it
out. However, no semiotic law of any significance was ever discovered, let
Communication
alone applied to linguistics. After the publication in 1957 of Noam
Chomsky's Syntactic Structures, linguistics took a new turn and did
undergo remarkable developments;4 but these owed nothing to semiotics.
As the structure of language became better understood, its sui generis
nature became more and more striking. The assumption that all systems of
signs should have similar structural properties became more and more
untenable. Without this assumption, however, the semiotic programme
makes little sense.
.
Saussure made a further prediction:
8
By studying rites, customs, etc. as signs, I believe that we shall throw

new light on the facts and point up the need for including them in a
science of semiology and explaining them by its laws. (1974: 17)
Here again, valiant attempts were made by anthropologists such as
Levi-Strauss or literary theorists such as Barthes to approach cultural or
artistic symbolism in semiotic terms. In the course of these attempts, they
certainly shed new light on the phenomena, and drew attention to many
interesting regularities; but they never came near to discovering an
underlying code in the strict sense: that is, a system of signal-message
pairs which would explain how myths and literary works succeed in
communicating more than their linguistic meaning, and how rites and
customs succeed in communicating at all.
This failure is instructive. What a better understanding of myth,
literature, ritual, etc., has shown is that these cultural phenomena do not,
in general, serve to convey precise and predictable messages. They focus
the attention of the audience in certain directions; they help to impose
some structure on experience. To that extent, some similarity of
representations between the artists or performers and the audience, and
hence some degree of communication, is achieved. However, this is a long
way from the identity of representations which coded communication is
designed to guarantee. It is not clear how the type of communication
involved in these cases could be explained in terms of the code model at all.
A semiotician might reply as follows. Granted that the best models we
have of human languages are generative grammars: since a generative
grammar just is a code which associates phonetic representations of
sentences to semantic representations of sentences, it follows that the code
model is applicable to verbal communication. Other forms of communication, say those involving Morse signals or traffic lights, are also
adequately described in terms of the code model. As for rites, customs and
the arts, although the semiotic approach is unable to deal with them yet,
there is no well-developed alternative approach either. Hence, the code
model is still the only available explanation of how communication is
possible at all.
Decoding and inference
We will try to show that this line of argument is invalid. It is true that a
language is a code which pairs phonetic and semantic representations of
sentences. However, there is a gap between the semantic representations
of sentences and the thoughts actually communicated by utterances. This
gap is filled not by more coding, but by inference. Moreover, there is an
alternative to the code model of communication. Communication has
been described as a process of inferential recognition of the communicator's intentions. We will try to show how this description can be improved
and made explanatory.5
2 Decoding and inference in verbal comprehension
As already mentioned, a generative grammar is a code which pairs

phonetic and semantic representations of sentences. Since an utterance can
generally be perceived as a realisation of the phonetic representation of a
single sentence (or in the case of phonetic ambiguity, two sentences), it is
reasonable to regard the phonetic representations of sentences as corresponding closely to the actual sounds of speech. By contrast, since most
sentences can be used to convey an infinite number of different thoughts,
the semantic representations of sentences cannot be regarded as corresponding very closely to thoughts. In constructing a general picture of
verbal communication, it is thus a legitimate idealisation (though phoneticians might not agree) to ignore the differences between phonetic
representations of sentences and acoustic realisations of utterances.
However, it is not legitimate to ignore the differences between the
semantic representations of sentences and the thoughts that utterances are
used to convey.
Crucial here is the difference between sentences and utterances. An
utterance has a variety of properties, both linguistic and non-linguistic. It
may contain the word 'shoe', or a reflexive pronoun, or a trisyllabic
adjective; it may be spoken on top of a bus, by someone with a heavy cold,
addressing a close friend. Generative grammars abstract out the purely
linguistic properties of utterances and describe a common linguistic
structure, the sentence, shared by a variety of utterances which differ only
in their non-linguistic properties. By definition, the semantic representation of a sentence, as assigned to it by a generative grammar, can take no
account of such non-linguistic properties as, for example, the time and
place of utterance, the identity of the speaker, the speaker's intentions, and
so on.
The semantic representation of a sentence deals with a sort of common
core of meaning shared by every utterance of it. However, different
utterances of the same sentence may differ in their interpretation; and
indeed they usually do. The study of the semantic representation of
10
Communication
sentences belongs to grammar; the study of the interpretation of

utterances belongs to what is now known as 'pragmatics'.6
To illustrate, consider sentences (1 )-(3):
(1) I'll come tomorrow.
(2) Bill is dll.
(3) Betsy's gift made her very happy.
A generative grammar cannot determine who 'I', 'Bill' and 'Betsy' refer to,
and which day 'tomorrow' picks out. It can only provide some very
general indications. It might state, for example, th<\;t 'I' always refers to the
speaker, that 'Bill' and 'Betsy' refer to people or other entities with those
names, and that 'tomorrow' picks out the day after the utterance. This is
not enough to determine which thought is expressed when sentences such
as (1)-(3) are uttered. For instance, if John says (1) on 25 March, it
expresses the thought that John will come on 26 March; if Ann says (1) on
30 November, it expresses the thought that Ann will come on 1
December. The grammar can say nothing about how the hearer, using
non-linguistic information, determines on a particular occasion what the
time of utterance actually is, who the speaker is, which Bill or Betsy the
speaker has in mind, etc., and hence which thought is actually being
expressed. These aspects of interpretation involve an interaction between
linguistic structure and non-linguistic information, only the former being
dealt with by the grammar.
Other aspects of the interpretation of (1)-(3) left unspecified by the
grammar are where the speaker of (1) is planning to come, by what criteria
Bill is tall (since, for instance, a tall dwarf is not a tall person), and in what
sense the ambiguous word 'gift' is to be taken. In every case, the grammar
can only help determine the possibilities of interpretation. How the hearer
sets about narrowing down and choosing among these possibilities is a
separate question. It is one that grammarians, but not pragmatists, can
ignore: an adequate theory of utterance interpretation must answer it.
Examples (1)-(3) show that as a result of referential indeterminacy such
as that of 'Bill', semantic ambiguity such as that of 'gift', and semantic
incompleteness such as that of 'tall', a single sentence, with a single
semantic representation, can express an unbounded range of thoughts.
There are still other factors widening the gap between sentence meaning
and utterance interpretation.
The same sentence, used to express the same thought, may sometimes
be used to present this thought as true, sometimes to suggest that it is not,
sometimes to wonder whether it is true, sometimes to ask the hearer to
make it true, and so on. Utterances are used not only to convey thoughts
but to reveal the speaker's attitude to, or relation to, the thought

expressed; in other words, they express 'propositional attitudes', perform
'speech-acts', or carry 'illocutionary force'.
To illustrate, consider sentences (4) and (5):
(4) You're leaving.
(5) What an honest fellow Joe is.
It makes a difference to the interpretation of (4) whether the speaker is
informing the hearer of a decision that he is to leave, making a guess and
asking him to confirm or deny it, or expressing outrage at the fact that he is
leaving. It makes a difference to the interpretation of (5) whether the
speaker is being sincere or ironical, making a literal claim or speaking
figuratively. Often, the linguistic structure of the utterance suggests a
particular attitude, as, for example, interrogative form most naturally
suggests that the utterance is a request for information. However, as
examples (4)-(5) show, the hearer is generally left a certain latitude, which
he must make up on the basis of non-linguistic information.
Moreover, an utterance which explicitly expresses one thought may
implicitly convey others. Whereas a thought that is explicitly expressed
must be in some kind of correspondence to the semantic representation of
the sentence uttered, those that are implicitly conveyed are under no such
constraint. Consider utterances (6) and (7):
(6) Do you know what time it is?
(7) Coffee would keep me awake.
The speaker of (6), while explicitly asking whether the hearer knows the
time, might be implicitly suggesting that it is time to go. The speaker of (7),
while making an explicit assertion about the effect of coffee, might be
implicitly refusing or forestalling an offer of coffee (or in other circumstances, implicitly soliciting or accepting such an offer).
Examples (1)-(7) show a variety of ways in which the semantic
representation of the sentence uttered may fall short of being a complete
interpretation of an utterance in context. As we have seen, code theorists
must show which code it is that makes verbal communication possible.
On closer examination, the claim that human languages, as described by
grammars which pair phonetic and semantic representations of sentences,
are codes of the required type is not borne out. This is not the end 9f the
code model of verbal communication, however. One might still assume
that the code involved is more complex than a grammar: rather than being
a grammar, it might merely contain a grammar as a sub-part.
To justify the code model of verbal communication, it would have to be
shown that the interpretation of utterances in context can be accounted for
by adding an extra pragmatic level of decoding to the linguistic level
12
Communication
provided by the grammar. Much recent work in pragmatics has assumed,
largely without question, that this can be done. Pragmatics has been
treated, on the analogy of phonology, syntax and semantics, as a code-like
mental device, underlying a distinct level of linguistic ability. It is widely
accepted that there are rules of pragmatic interpretation much as there are
rules of semantic interpretation, and that these rules form a system which
is a supplement to a grammar as traditionally understood.
There are certainly pragmatic phenomena which lend themselves to this
sort of approach. For example, a pragmatic device might contain rules of
interpretation such as (8) and (9):
(8) Substitute for 'I' a reference to the speaker.
(9) Substitute for 'tomorrow' a reference to the day after the utterance.
Imagine a hearer equipped with such rules and able to recognise that the
speaker of (1) is Ann and the date of utterance is 30 November. He could
automatically interpret utterance (1) as conveying the thought in (10):
(1) I'll come tomorrow.
(10) Ann will come on 1 December.
However, most aspects of utterance interpretation cannot be handled so

easily. Consider (11) and (12):
(11) He's got egg on his tie.
(12) That's interesting.
It presumably follows from the grammar of English that the referent of
'he' must be male and the referent of 'that' must be non-human. However,
(11) and (12) are unlike (1) in that on virtually every occasion of utterance,
there is more than one referent meeting these conditions. The assignment
of actual referents in these cases must clearly involve something much
more complicated than rules (8) and (9).
To substantiate the code model of verbal communication, it would have
to be shown that every case of reference assignment can be dealt with by
rules which automatically integrate properties of the context with
semantic properties of the utterance. It would also have to be shown that
disambiguation, the recovery of propositional attitudes, figurative interpretations and implicit import can be handled along similar lines. Nothing
approaching such a demonstration has ever been given.
While still assuming that the code model provides the framework for a
general theory of communication, and hence for a theory of verbal
communication, most pragmatists have described comprehension as an
inferential process. Inferential and decoding processes are quite different.
An inferential process starts from a set of premises and results in a set of
conclusions which follow logically from, or are at least warranted by, the
13
premises. A decoding pr?ces~ starts. from a sigr;tal and results in the
recovery of a message whICh IS assocIated to the sIgnal by an underlying
code. In general, conclusions are not associated to their premises by a
code, and signals do not warrant the messages they convey.
To illustrate the difference between coding and inferential processes,
consider (13)-(15):
(13) (a) Either Mary is early or Bob is late.
(b) Bob is never late.
(14) [m;;)ri: IZ3 :li:]
(15) Mary is early.
That Mary is early, i.e. (15), can be either inferred from the premises in
(13) or decoded from the phonetic signal in (14), but the converse is not
true: (15) can be neither decoded from (13) nor inferred from (14). It
cannot be decoded from (13) because there is no code identifying (13) as a
signal and (15) as its associated message. It cannot be inferred from (14)
because signals do not by themselves warrant the messages they encode
(otherwise any absurdity could be transformed into a warranted assumption merely by uttering it).
The view that utterance interpretation is a largely inferential process
squares well with ordinary experience. Consider (16)-(18), for instance:
(16) Jones has bought the Times.
(17) Jones has bought a copy of the Times.
(18) Jones has bought the press enterprise which publishes the Times.
Sentence (16) is ambiguous, and can be understood as conveying either

(17) or (18). Ordinary hearers in ordinary circumstances have no trouble
choosing one of these two meanings, usually without even realising that
they have made a choice. When the ambiguity is pointed out and they are
asked to explain how they know which interpretation is correct, they
generally offer something that looks like a truncated logical argument: the
speaker must have intended this interpretation rather than that, because
this is the only interpretation that is true; or the only one that gives the
required information; or the only one that makes sense.
For instance, hearers asked why they understood 'Jones has bought the
Times' to mean 'Jones has bought a copy of the Times' rather than 'Jones
has bought the press enterprise which publishes the Times' might answer:
'because the other interpretation could not be true', or 'because the
question was whether I should buy a copy of the Times myself'. The
assumption behind these truncated arguments is that speakers set
themselves certain standards, of truthfulness, informativeness, comprehensibility, and so on, and only try to communicate information that
meets the standards set. As long as speakers systematically observe the
14
Communication
standards, and hearers systematically expect them to, a whole range of

linguistically possible interpretations for any given utterance can be
inferentially dismissed, and the task of communication and comprehension becomes accordingly easier. The same types of truncated argument,
based on implicit standards, are invoked by hearers to justify their
interpretation of referential expressions, illocutionary force, figures of
speech and implicit import.
Modern pragmatists, inspired by the work of Grice/ have tried to
describe these implicit standards of verbal communication more explicitly
and show how they are used in comprehension. The mental processes
involved have not been described in any detail, but everybody agrees that
they are inferential. As we have said, inferential processes are quite
different from decoding processes. Does it follow that pragmatists who
hold to the code model, and yet describe comprehension in inferential
terms, are being inconsistent? Not necessarily: an inferential process can
be used as part of a decoding process.
Let us use an artificial example to show how inference can double as
decoding. Imagine two partners who know (when nobody else around
them knows) that (19) is true, who want to let one another know whether
(20) is true, and who do not want bystanders to benefit from the
information:
(19) Bob is in Miami.
(20) The speaker will leave the party.
They can use the standard inference rule (21) as a decoding rule, treat
utterances (22) and (23) as signals, and thus convey by use of these signals
messages (24) and (25) respectively:
(21) Premises: If P then Q

P
Conclusion: Q
(22) If Bob is in Miami, I'll leave the party.
(23) If Bob is in Miami, I won't leave the party.
(24) The speaker will leave the party.
(25) The speaker will not leave the party.
In this example, we have an inferential process simultaneously functioning
as a decoding process. However, for this to be possible several conditions
have to be fulfilled: first, speaker and hearer must share the tacit premise
(19); second, they must share the inference rule (21); and third, they must
use that premise and that rule to the exclusion of any other tacit premise or
inference rule at their disposal. Otherwise, the signal will not be properly
decoded.
15
Do speakers and hearers in ordinary verbal communication generally

achieve a similar parallelism of premises and inference rules? If not, the
inferential processes involved in verbal comprehension cannot qualify as
decoding processes. To defe.nd the code model of verbal communication,
it must be shown, then, how speaker and hearer can come to have not only
a common language, but also common sets of premises, to which they
apply identical inference rules in parallel ways.
For language, the demonstration is fairly straightforward. The evidence
suggests that speakers with quite different linguistic histories may end up
with very similar grammars. Any number of different examples will do to
illustrate a particular aspect of linguistic structure - say, the relative clause
- so that it does not much matter which utterances of the language the
child actually hears. It is also clear that after a certain point, the structure
of the language has essentially been mastered, so that as new utterances are
encountered, the grammar of an adult speaker will hardly change at all.
The requirement of a common language thus presents no real difficulty for
the code model.
Although the question of inference rules has not been dealt with in the
pragmatic literature, it is arguable that the development of inferential
abilities is similar in relevant respects to that of linguistic abilities. That is,
any application of an inference rule will give grounds for its adoption.
Thus different experiences with inferential processes may nevertheless
converge on the same logical system. A more serious problem is that
logical systems, as described by logicians, allow infinitely many different
conclusions to be derived from the same premises. How, then, is the
hearer to infer just those conclusions intended by the speaker? A solution
to this problem will be proposed in the next chapters.
However, as we will show in the next section, the claim that speaker and
hearer can and do restrict themselves to a set of common premises is much
harder to maintain.
3 The mutual-knowledge hypothesis
The set of premises used in interpreting an utterance (apart from the

premise that the utterance in question has been produced) constitutes
what is generally known as the context. A context is a psychological
construct, a subset of the hearer's assumptions about the world. It is these
assumptions, of course, rather than the actual state of the world, that affect
the interpretation of an utterance. A context in this sense is not limited to
information about the immediate physical environment or the immediately preceding utterances: expectations about the future, scientific hypoth-
16
Communication
eses or religious beliefs, anecdotal memories, general cultural assump~ions, beliefs about the mental state of the speaker, may all play a, role in
mterpretation.
While it is clear that members of the same linguistic community
converge on the same language, and plausible that they converge on the
same inferential abilities, the same is not true of their assumptions about
the world. True, all humans are constrained by their species-specific
cognitive abilities in developing their representation of the world, and all
members of the same cultural group share a number of experiences,
teachings and views. However, beyond this common framework, individuals tend to be highly idiosyncratic. Differences in life history
necessarily lead to differences in memorised information. Moreover, it has
been repeatedly shown that two people witnessing the same event - even a
salient and highly memorable event like a car accident - may construct
dramatically different representations of it, disagreeing not just on their
interpretation of it, but in their memory of the basic physical facts. 8 While
grammars neutralise the differences between dissimilar experiences,
cognition and memory superimpose differences even on common experiences.
Grammars and inferential abilities stabilise after a learning period and
remain unchanged from one utterance or inference to the next. By
contrast, each new experience adds to the range of potential contexts. It
does so crucially in utterance interpretation, since the context used in
interpreting a given utterance generally contains information derived
from immediately preceding utterances. Each new utterance, while
drawing on the same grammar and the same inferential abilities as previous
utterances, requires a rather different context. A central problem for
pragmatic theory is to describe how, for any given utterance, the hearer
finds a context which enables him to understand it adequately.
A speaker who intends an utterance to be interpreted in a particular way
must also expect the hearer to be able to supply a context which allows that
interpretation to be recovered. A mismatch between the context envisaged
by the speaker and the one actually used by the hearer may result in a
misunderstanding. Suppose, for example, that the speaker of (7) wants to
stay awake, and therefore wants to accept his host's offer of coffee,
whereas the host assumes that the speaker does not want to stay awake,
and thus interprets (7) as a refusal:
(7) Coffee would keep me awake.
Clearly, this difference between actual and envisaged contexts will lead to
a misunderstanding. Of course such misunderstandings do occur. They
are not attributable to noise in the acoustic channel. The question is
whether they happen because the mechanisms of verbal communication
17
are sometimes improperly applied, or because these mechanisms at best

make successful communication probable, but do not guarantee it. We
will pursue this second alternative. Most pragmatists opt for the first: they
try to describe a failsafe mechanism which, when properly applied and not
disrupted by noise, would guarantee successful communication.
The only way to make sure that misunderstandings such as the one
described above could not arise would be to make sure that the context
actually used by the hearer was always identical to the one envisaged by
the speaker. How can this be done? Since any two people are sure to share
at least a few assumptions about the world, they should use only these
shared assumptions. However, this cannot be the whole answer, since it
immediately raises a new question: how are the speaker and hearer to
distinguish the assumptions they share from those they do not share? For
that, they must make second-order assumptions about which first-order
assumptions they share; but then they had better make sure that they share
these second-order assumptions, and that calls for third-order assumptions. Some pragmatists stop here (e.g. Bach and Harnish 1979) and
consider it of no practical importance that in principle, as noticed by
others (Schiffer 1972; Clark and Marshall 1981), the same problem arises
for third-order assumptions, calling for fourth-order assumptions, and so
on indefinitely.
Consider a relevant example from the literature on reference assignment:
On Wednesday morning Ann and Bob read the early edition of the
newspaper, and they discuss the fact that it says that A Day at the
Races is showing that night at the Roxy. When the late edition
arrives, Bob reads the movie section, notes that the film has been
corrected to Monkey Business, and circles it with his red pen. Later,
Ann picks up the late edition, notes the correction, and recognizes
Bob's circle around it. She also realizes that Bob has no way of
knowing that she has seen the late edition. Later that day Ann sees
Bob and asks, 'Have you ever seen the movie showing at the Roxy
tonight?' (Clark and Marshall1981: 13)
The question is, which film should Bob take Ann to be referring to? As
Clark and Marshall point out, although Ann and Bob both know that the
film showing at the Roxy is Monkey Business, and Ann knows that Bob
knows that it is, this degree of shared knowledge is not enough to
guarantee successful communication. Bob might reason that although he
knows that the film actually showing is Monkey Business, Ann might still
thinkit is A Day at the Races, and be referring to that. Or he might decide
that she must have seen the marked correction, have realised that he knows
18
Communication
the film is Monkey Business, and be referring to that. Or perhaps he might

think that though she must have seen the correction, she will realise that he
has no way of knowing that she has, so she will in fact be referring to A
Day at the Races. Or maybe she has seen the correction and expects him to
realise that she has seen it, but is not sure he will realise that she will realise
that he will realise that she has seen it; and so on ad infinitum.
Clark and Marshall conclude that the only way to guarantee successful
communication is for Ann not only to know what the film showing at the
Roxy actually is, but to know that Bob knows what it is, and that Bob
knows that she knows what it is, and that he knows
that she knows that he
@
knows what it is, and so on indefinitely. Similarly, Bob must not only
know what the film showing at the Roxy actually is, but know that Ann
knows what it is, and that she knows that he knows what it is, and that she
knows that he knows that she knows what it is, and so on indefinitely.
Knowledge of this infinitely regressive sort was first identified by Lewis
(1969) as common knowledge, and by Schiffer (1972) as mutual
knowledge. 9 The argument is that if the hearer is to be sure of recovering
the correct interpretation, the one intended by the speaker, every item of
contextual information used in interpreting the utterance must be not only
known by the speaker and hearer, but mutually known.
Within the framework of the code model, mutual knowledge is a
necessity. If the only way to communicate a message is by encoding and
decoding it, and if inference plays a role in verbal communication, then the
context in which an utterance is understood must be strictly limited to
mutual knowledge; otherwise inference cannot function as an effective
aspect of decoding. But as virtually everyone who has touched on the topic
has noticed, it is hard to see how the requirement of mutual knowledge
could ever be built into a psychologically adequate account of utterance
production and comprehension. Someone who adopts this hypothesis is
thus inevitably forced to the conclusion that when human beings try to
communicate with each other, they are aiming at something they can
never, in fact, achieve.
If mutual knowledge is necessary for communication, the question that
immediately arises is how its existence can be established. How exactly do
the speaker and hearer distinguish between knowledge that they merely
share, and knowledge that is genuinely mutual? To establish this
distinction, they would have, in principle, to perform an infinite series of
checks, which clearly cannot be done in the amount of time it takes to
produce and understand an utterance. Hence, even if they try to restrict
themselves to what is mutually known, there is no guarantee that they will
succeed.
Many pragmatists have accepted this conclusion and argued that mutual
knowledge is not a reality but 'an ideal people strive for because they ... want
19
to avoid misunderstanding whenever possible' (Clark and Marshall1981:

27). Now while it is true that people sometimes go to great lengths to avoid
misunderstanding, such efforts are the exception and not the rule. In legal
proceedings, for instance, there really is a serious attempt to establish
mutual knowledge among all the parties concerned: all laws and
precedents are made public, all legitimate evidence is recorded, and only
legitimate evidence can be considered, so that there is indeed a restricted
domain of mutual knowledge on which all parties may call, and within
which they must remain. There is no evidence of any such concern in
normal conversation, however serious or formal it is. All sorts of risks are
taken, assumptions and guesses made. There is no indication that any
particular striving after mutual knowledge goes on.
Enormous energy has been spent on trying to develop an empirically
defensible approximation to the mutual knowledge requirement. It has
been argued that in certain circumstances, speaker and hearer are justified
in assuming that they have mutual knowledge, even though its existence
cannot be conclusively established. For example, if two people can see
each other looking at the same thing, they have grounds for assuming
mutual knowledge of its presence. If some information has been verbally
given in their joint presence, they are justified in assuming mutual
knowledge of it. If some fact is known to all members of a community,
twO people who think they recognise each other as members of that
community have grounds for assuming mutual knowledge of that fact. In
none of these cases, though, can there be any certainty of mutual
knowledge. People may look at the same object and yet identify it
differently; they may impose different interpretations on information that
they are jointly given; they may fail to recognise facts. In all these cases,
the individual would be wrong in assuming mutual knowledge.
There is a paradox here. Since the assumption of mutual knowledge may
always be mistaken, the mutual-knowledge hypothesis cannot deliver the
guarantees it was set up to provide. If Bob may be mistaken in assuming
that he and Ann have mutual knowledge of the fact that the film playing at
the Roxy is Monkey Business, he cannot be sure of having correctly
understood which film she is referring to. Bob's painstaking but
inconclusive attempt at ascertaining mutual knowledge does not really
protect him from the risk of misunderstanding. So why go to all that
trouble?
There is yet another paradox in the idea that speaker and hearer might
reasonably come to assume, but with something less than certainty, that
they have mutual knowledge of some fact. By the very definition of
mutual knowledge, people who share mutual knowledge know that they
do. If you do not know that you have mutual knowledge (of some fact,
with someone), then you do not have it. Mutual knowledge must be
Communication
certain, or else it does not exist; and since it can never be certain it can
never eXIst.
The apparent fallback position for the code theorist would be to replace
the requirement of mutual knowledge by that of mutual probabilistic
assumptions. This more realistic proposal raises an obvious problem. In
general, the higher the order of the assumptions involved in such a scheme,
the less likely they are to be true. Bob may know for a fact that Monkey
Business is the film playing tonight; in the absence of compelling evidence,
he should feel less certain that Ann assumes that he knows it, and even less
certain that she assumes that he assumes that she assumes that he knows it,
and so on. The assumption of mutuality itself, which is the highest ordered
one, will have the weakest probability. How, then, could restricting the
context to mutual assumptions ensure the identity or near-identity of
premises which the code model requires?
Another problem with the mutual-knowledge hypothesis is that even if
it defines a class of potential contexts for use in utterance interpretation, it
says nothing about how an actual context is selected, nor about the role of
context in comprehension. Take the following utterance:
20
(26) The door's open.
Speaker and hearer might have shared knowledge of hundreds of different

doors; the mutual-knowledge requirement does nothing to explain how
the choice of an actual referent is made.
Bach and Harnish (1979: 93) spend some time justifying their particular
version of the mutual-knowledge hypothesis, but add that their pragmatic
theory says little about 'the specific strategy the hearer uses to identify a
particular communicative intent. It gives no indication of how certain
mutual beliefs are activated or otherwise picked out as relevant, much less
how the correct identification is made.' But in that case, the adoption of
the mutual-knowledge hypothesis is just whistling in the dark. Until we
know something about how contexts are actually selected and used in
utterance interpretation, the belief that they must be restricted to mutual
knowledge has no justification apart from the fact that it follows from the
code model. 10
Pragmatists have no positive argument that individuals engaging in
verbal communication can and do distinguish mutual from non-mutual
knowledge. Their only argument is a negative one: if mutual knowledge
does not exist in the form required by the code model of verbal
communication, then the code model is wrong. Since they see the code
model as the only possible explanation of communication, they cling to
the mutual-knowledge hypothesis.
Instead of adopting the code model, seeing that it commits us to the
mutual-knowledge hypothesis, and then having to worry about how this
Grice, 'meaning' and communication
21
hypothesis can be empirically justified, we want to approach things the

other way around. We see the mutual-knowledge hypothesis as untenable. We conclude, therefore, that the code theory must be wrong, and
that we had better worry about possible alternatives.
4 Grice's approach to 'meaning' and communication

In 1957, Paul Grice published an article, 'Meaning', which has been the
object of a great many controversies, interpretations and revisions.l1 In
this article, Grice proposed the following analysis of what it is for an
individual 5 to mean something by an utterance x (where 'utterance' is to
be understood as referring not just to linguistic utterances but to any form
of communicative behaviour):
'CS] meant something by x' is (roughly) equivalent to '[5] intended

the utterance of x to produce some effect in an audience by means of
the recognition of this intention'. (Grice 1957/1971: 58)
Strawson's reformulation of this analysis (Strawson 1964a/1971: 155; see
also Schiffer 1972: 11) separates out the three sub-intentions involved. To
mean something by x, 5 must intend
(27) (a) S's utterance of x to produce a certain response r in a certain
audience A;
(b) A to recognise 5's intention (a);
(c) A's recognition of 5's intention (a) to function as at least part of
A's reason for A's response r.
This analysis can be developed in two ways. Grice himself used it as the
point of departure for a theory of 'meaning', trying to go from the analysis
of 'speaker's meaning' towards such traditional semantic concerns as the
analysis of 'sentence meaning' and 'word meaning'. For reasons which
should become apparent, we doubt that very much can be achieved in this
direction. However, Grice's analysis can also be used as the point of
departure for an inferential model of communication, and this is how we
propose to take it. In the rest of this section we will show how this analysis
applies to the description of communication. In the next three sections we
will consider some of the objections and reformulations which have been
proposed. Finally, in the last five sections of this chapter, we will develop
our own model.
There are situations in which the mere fact that an intention is
recognised may lead to its fulfilment. Suppose that Mary intends to please
22
Communication
Peter. If Peter becomes aware of her intention to please him, this may in
itself be enough to please him. Similarly, when the inmates of a prison
recognise their warder's intention to make them fear him, this may be
enough in itself to make them fear him. There is one type of intention for
which this possibility, rather than being exceptional, is regularly exploited: intentions to inform are quite generally fulfilled by being made
recognisable.
Suppose that Mary intends to inform Peter of the fact that she has a sore
throat. All she has to do is let him hear her hoarse voice, thus providing
him with salient and conclusive evidence that she has a sore throat. Here,
Mary's intention can be fulfilled whether or not Peter is aware of it: he
could realise that she has a sore throat without also realising that she
intends him to realise that she has one. Suppose now that Mary intends, on
2 June, to inform Peter (truly or falsely) that she had a sore throat on the
previous Christmas Eve. This time she is unlikely to be able to produce
direct evidence of her past sore throat. What she can do, though, is give
him direct evidence, not of her past sore throat, but of her present
intention to inform him of it. How can she do this, and what good will it
do? One way she can do it is by uttering (28), and the good it will do is to
give Peter indirect, but nevertheless strong, evidence that she had a sore
throat on the previous Christmas Eve:
(28) I had a sore throat on Christmas Eve.
In our first example, Mary's hoarse voice is most likely to have been
caused by her sore throat. The fact that she has spoken hoarsely is thus
direct evidence for the assumption that she has a sore throat. Mary's
utterance of (28) on 2 June is not directly caused by her having had a sore
throat on the previous Christmas Eve. Hence her utterance is not direct
evidence for the assumption that she had a sore throat on the previous
Christmas Eve. However, her utterance is directly caused by her present
intentions. Although she might have had various intentions in uttering
(28), it is most likely that she intended to inform Peter that she had a sore
throat on the previous Christmas Eve. This makes Mary's utterance direct
evidence of her present intention to inform Peter of her past sore throat.
Suppose now that Peter assumes that Mary is sincere and is likely to
know whether or not she had a sore throat on the previous Christmas Eve.
Then for Peter, the fact that Mary intends to inform him that she had a sore
throat on that date provides conclusive evidence that she had. In these
quite ordinary conditions, Mary's intention to inform Peter of her past
sore throat can be fulfilled by making Peter recognise her intention. This is
not an exceptional way of fulfilling an intention to inform an audience. Let
us assume that it is precisely how Mary intends to have her intention
Grice, 'meaning' and communication
23
fulfilled. Then she does have all three sub-intentions of the GriceStraws on definition (27), as shown in (29):
(29) Mary intends
(a) her utterance (28) to produce in Peter the belief that she had a sore
throat the previous Christmas Eve;
(b) Peter to recognize her intention (a);
(c) Peter's recognition of her intention (a) to function as at least part
of his reason for his belief.
Mary's intentions in this example are quite similar in structure to those we

all have when we communicate, verbally or otherwise.
We have shown two different ways of conveying information. One way
is to provide direct evidence for the information to be conveyed. This
should not be regarded as a form of communication: any state of affairs
provides direct evidence for a variety of assumptions without necessarily
communicating those assumptions in any interesting sense. Another way
of conveying information is to provide direct evidence of one's intention
to convey it. The first method can only be used with information for
which direct evidence can be provided. The second method can be used
with any information at all, as long as direct evidence of the communicator's intentions can be provided. This second method is clearly a form of
communication; we will call it, for the time being, inferential communication (and, in section 10, ostensive-inferential communication): it is
inferential in that the audience infers the communicator's intention from
evidence provided for this precise purpose.
The description of communication in terms of intentions and inferences
is, in a way, commonsensical. We are all speakers and hearers. As speakers,
we intend our hearers to recognise our intention to inform them of some
state of affairs. As hearers, we try to recognise what it is that the speaker
intends to inform us of. Hearers are interested in the meaning of the
sentence uttered only insofar as it provides evidence about what the
speaker means. Communication is successful not when hearers recognise
the linguistic meaning of the utterance, but when they infer the speaker's
'meaning' from it. This is shown by the following easily verifiable
observation: when hearers realise that the speaker has misused a word or
made a slip of the tongue, they generally discount the wrong meaning. The
meaning they discount, however, need not be ill-formed or undecodable;
it is 'wrong' only in that it provides misleading evidence about the
speaker's intentions.
From a psychological point of view, the description of communication
in terms of intentions and inferences also makes good sense. The
attribution of intentions to others is a characteristic feature of human
24
Communication
cognition and interaction. Humans typically conceptualise human and
animal behaviour, not in terms of its physical features, but in terms of its
underlying intentions. For instance, an ordinary-language concept such as
give, take, attack or defend applies to various forms of behaviour which
do not fall under any characteristic physical description, and have in
common only the kind of intention which governs them. Human
interaction is largely determined by the conceptualisation of behaviour in
intentional rather than physical terms. The idea that communication
exploits this ability of humans to attribute intentions to each other should
be quite intelligible, and even appealing, to cognitive and social psychologists.
So it seems that we all know - semioticians included - that communication involves the publication and recognition of intentions. Yet until
Grice, the significance of this truism was generally ignored;12 attempts to
describe and explain communication continued to be based on one form
or another of the code model. Grice's original idea, as presented in his
1957 paper, can thus be seen as an attempt to rehabilitate a commonsense
view of communication and spell it out in theoretically acceptable terms.
However, the elaboration of this idea in the work of Grice himself,
Strawson, Searle, Schiffer and others has often taken the form of a move
away from common sense, away from psychological plausibility, and
back to the code model. This unfortunate development resulted from the
discovery of part spurious, part genuine problems with Grice's original
formulation.
5 Should the code model and the inferential model be amalgamated?
We have now looked at two models of communication. According to the

code model, communication is achieved by encoding and decoding
messages. According to the inferential model, communication is achieved
by the communicator providing evidence of herB intentions and the
audience inferring her intentions from the evidence. Several questions
come to mind. Are these two different models of the same thing? If so,
must we choose between them, or can they be amalgamated in some way?
Or are they, as we have hinted, models of two quite different things? If so,
how are these things related?
Most theorists see communication as a unitary phenomenon, to be
described by a single model. The code model is very well entrenched in the
Western scholarly tradition. The inferential model appeals to common
sense. When an appealing new approach is put forward, the temptation is
to treat it not as an alternative to the old approach but as an elaboration of
it. This is what most pragmatists have done, almost unconsciously, with
The code and inferential models
25
Grice's analysis. John Searle at least takes the trouble to justify this
reaction. i4 He claims that Grice's analysis
fails to account for the extent to which meaning can be a matter of
rules and conventions. This account of meaning does not show the
connection between one's meaning something by what one says, and
what that which one says actually means in the language. (Searlt;
1969: 43)
Searle wants to improve on Grice's account by showing the connection
between speaker's meaning and -linguistic meaning. His first step is to
restrict the application of this account to the domain of 'literal meaning'.
This he defines in terms of the speaker's intentions, including the intention
to have her intentions recognised, but adds a rider: the speaker should
intend the hearer to recognise her intentions 'in virtue of his knowledge of
the rules for the sentence uttered' (Searle 1969: 48). In other words, the
speaker should intend the hearer to understand her by decoding her
utterance.
This reduces Grice's analysis to a commonsense amendment of the code
model. The code model is reintroduced as the basic explanation of
communication, but in the case of human communication, the message
th~t is encoded and decoded is regarded as a communicator's intention. If
Searle's revision is justified, then Grice's analysis is not a genuine
alternative to the code model after all.
Grice's greatest originality was not to suggest that human communication involves the recognition of intentions. That much, as already pointed
out, is common sense. It was to suggest that this characterisation is
sufficient: as long as there is some way of recognising the communicator's
intentions, then communication is possible. Recognition of intentions is
an ordinary human cognitive endeavour. If Grice is right, the inferential
abilities that humans ordinarily use in attributing intentions to each other
should make communication possible even in the absence of a code. And
of course it is possible.
For example, Peter asks Mary,
(30) How are you feeling today?
Mary responds by pulling a bottle of aspirin out of her bag and showing it
to him. Her behaviour is not coded: there is no rule or convention which
says that displaying a bottle of aspirin means that one is not feeling well.
Similarly, her behaviour affords only the weakest kind of direct evidence
about her feelings: maybe she always carries a bottle of aspirin in her bag.
On the other hand, it is strong direct evidence of her intention to inform
26
Communication
Peter that she does not feel well. Because her behaviour enables Peter to
recognise her intention, Mary successfully communicates with him, and
does so without the use of any code. 15
Even Searle does not deny the existence of purely inferential communication. However, he insists that it is rare, and that most human
communication crucially involves the use of a language or code:
Some very simple sorts of illocutionary acts can indeed be performed
apart from any use of conventional devices at all, simply by getting
the audience to recognize certain of one's ip.tentions in behaving in a
certain way ... One can in certain special circumstances 'request'
someone to leave the room without employing any conventions, but
unless one has a language one cannot request of someone that he,
e.g., undertake a research project on the problem of diagnosing and
treating mononucleosis in undergraduates in American universities.
(Searle 1969: 38)
It may be true that most human communication involves the use of

language, that cases of communication clearly achieved without the use of
a code are rare, and that the thoughts so communicated tend to be rather
simple. But the very existence of such cases is incompatible with the code
model. On the other hand, it is predicted by the inferential model. Searle's
dismissal of these cases as unimportant misses the point. They may be
unimportant as examples of human interaction, but they are important as
evidence for or against theories.
Since purely inferential communication exists, the inferential model is
adequate by itself to account for at least some forms of communication.
On the other hand, there is no doubt that most cases of communication
involve the use of a code. Someone who takes the strong view that all
human communication must be accounted for in inferential terms is then
faced with the task of redescribing coding and decoding in inferential
terms. Here is how it might be done. Regard a code as a set of conventions
(in the sense of Lewis 1969) shared by all participants in the communication process. Members of the audience use their knowledge of these
conventions on the one hand, and their knowledge of the signal and of the
context on the other, to infer the message. This is a reasonably good
description of what often happens when artificial codes are devised and
used.
For example, Romeo and Juliet agree between them that a white
kerchief tied to the rail of her balcony means that he can come up. Romeo
sees the white kerchief, uses as a premise his knowledge of the convention
they have devised, i.e. his knowledge that a white kerchief means that he
can come up, and indeed infers that he can come up. When this account is
The code and inferential models
27
generalis~d, all ~ecoding, lir:g~isti~ decoding included, is s~e~ as an

ordinary mferentIal process dIstmgUIshed only by the fact that It mvolves
premises based on knowledge of linguistic conventions.
We believe, and will argue in a later chapter, that the strong inferential
theory of communication is empirically inadequate. There are codingdecoding processes, and there are inferential processes, and the two types
of process are essentially distinct (even though, under rather artificial
conditions, inference can mimic decoding, or decoding can mimic
inference). A variety of species, from bees to humans, have codes which
are to a greater or lesser extent genetically determined. These differ from
inferential systems in two main respects: first, the representations they
relate need not be conceptual, and second, the rules relating these
representations need not be inferential. Human natural languages are cases
in point. If we are right, then linguistic knowledge does not contribute to
the comprehension process in the way described above: by providing
premises for inference. 16
We maintain, then, that there are at least two different modes of
communication: the coding-decoding mode and the inferential mode. If
we are right, from the fact that a particular communication process
involves the use of a code, it does not follow that the whole process must
be accounted for in terms of the code model. Complex forms of
communication can combine both modes. Inferential communication, for
example, might involve the use of coded signals which fall short of
encoding the communicator's intentions and merely provide incomplete
evidence about them. It becomes an empirical question whether the code
model can provide a full account of a given communication process. It is
not enough to show that a code is being used; one must also be able to
show that what is communicated is actually being encoded and decoded.
Otherwise, all that can be reasonably maintained is that the use of a code
plays some role in this particular communication process, without
perhaps wholly explaining it.
Verbal communication is a complex form of communication. Linguistic
coding and decoding is involved, but the linguistic meaning of an uttered
sentence falls short of encoding what the speaker means: it merely helps
the audience infer what she means. The output of decoding is correctly
treated by the audience as a piece of evidence about the communicator's
intentions. In other words, a coding-decoding process is subservient to a
Gricean inferential process.
Searle saw the fact that almost all human communication involves the
use of codes as an objection to Grice's analysis. However, this fact is easy
to explain on the assumption that the code and inferential modes of
communication can combine. People who are in a position to communicate with one another usually share a language (and various minor codes);
28
Communication
as a result, they can produce much subtler and stronger evidence about
their intentions than they could in the absence of a shared code. They are
unlikely, then, to go to the trouble of communicating inferentially
without these powerful tools, just as modern humans are unlikely to go to
the trouble of making fire without matches or lighters. Still, just as no one
would want to define fire as necessarily produced by the use of matches or
lighters, it would be unreasonable to define communication as necessarily
achieved by the use of codes.
The reduction of Grice's analysis to an amendment of the code model
destroys not just its originality, but also many of i~s empirical implications
and justifications. The elevation of the inferential model into a general
theory of communication ignores the diversity of forms of communication, and the psychological evidence that much decoding is noninferential (to be discussed in chapter 4).
6 Problems of definition
Most discussions of Grice's 1957 article have had to do with the definition
of 'meaning' or 'communication' and have been highly philosophical. In
this section, we will single out two genuinely empirical issues for
discussion. Our aim is simply to highlight these relevant issues, not to .
write a history or an evaluation of the surrounding debates.
Grice characterises 'meaning' in terms of a communicator's intentions.
Conversely, an act of communication (in an appropriately restricted sense
of the term) might be characterised as one that fulfils these Gricean
intentions. However, as Searle (1969: 46-8; 1971: 8-9) points out, a
communicator can mean something, and successfully communicate it,
without all these Gricean intentions being fulfilled. Recall Strawson's
reformulation (27) of Grice's analysis. To mean something by an utterance
x, an individual S must intend
(27) (a) S's utterance of x to produce a certain response r in a certain
audience A;
(b) A to recognise S's intention (a);
(c) A's recognition of S's intention (a) to function as at least part of A's
reason for A's response r.
Now it is easy to see that once intention (b) is fulfilled, the
communicator has succeeded in communicating what she meant, whether
or not intentions (a) and (c) are also fulfilled. For example, when Mary
utters (28), her specific intention (29a) is to produce in Peter the belief that
she had a sore throat on the previous Christmas Eve. Suppose Peter
recognises this intention, but does not believe Mary. Then only her
Problems of definition
29
intention (29b) is fulfilled; intentions (29a) and (29c) are not. Nonetheless,
although Mary has failed to convince Peter, she. has succeeded in
communicating to him what she meant.
Since communication can succeed without intention (27a) being
fulfilled, intention (27a) is not an intention to communicate at all. It is
better described as an intention to inform, or as we will call it, an
informative intentionY The. true communicative intention is intention
(27b): that is, the intention to have one's informative intention recognised.
What about intention (27c): that the recognition by the audience of the
communicator's intention (27a) shall function as at least part of the
audience's reason for fulfilling intention (27a)? By definition, intention
(27c) cannot be fulfilled when the informative intention (27a) is not. Since
the fulfilment of (27a) is not necessary for successful communication, the
fulfilment of (27c) cannot be necessary either. What Grice has convincingly shown is that the recognition of an informative intention can lead to its
fulfilment. Very often, it is because this possibility exists that the
communicator engages in communication at all. However to turn this
possibility into a definitional necessity requires some justification. For the
time being, we will drop intention (27c) from the characterisation of
inferential communication without further discussion, and re-examine
Grice's motivations on this point in section 10. 18
We are now almost ready to propose a modified version of Grice's
analysis, highlighting the difference between the informative and communicative intentions. However, we must first get rid of a confusing
terminological idiosyncrasy. Grice and Strawson use the term 'utterance'
to refer not just to linguistic utterances, or even to coded utterances, but to
any modification of the physical environment designed by a communicator to be perceived by an audience and used as evidence of the
communicator's intentions. This usage seems to us to introduce a bias into
the identification of communicative behaviour. It encourages the view
that utterances in the usual linguistic sense can be taken as the paradigm of
communicative behaviour in general. Psychologists use the term 'stimulus' for any modification of the physical environment designed to be
perceived. We will do the same. An utterance in the usual sense is, of
course, a special case of a stimulus. Let us say, then, that communication
involves producing a certain stimulus intending theFeby
(31) Informative intention: to inform the audience of something;
Communicative intention: to inform the audience of one's
informative intention.
Note that the communicative intention is itself a second-order informative intention: the communicative intention is fulfilled once the first-order
informative intention is recognised. In ordinary situations, if all goes well,
30
Communication
the recogmtion of the informative intention will itself lead to the

fulfilment of that intention, so that as a result of an act of communication,
both the communicative and the informative intention will be fulfilled.
However, a communicative intention can be fulfilled without the
corresponding informative intention being fulfilled. Hence our reformulation is not open to the objections to Grice's and Strawson's
formulation that we have so far considered.
While Grice's conditions on communication are too restrictive in some
respects, in others they are not restrictive enough. One tends to think of
communication as something done overtly: either@your behaviour makes
it clear that you are communicating, or else you are not truly communicating at all. In other words, communication should be distinguished from
covert forms of information transmission.
Suppose, for instance, that Mary wants Peter to mend her broken
hair-drier, but does not want to ask him openly. What she does is begin to
take her hair-drier to pieces and leave the pieces lying around as if she were
in the process of mending it. She does not expect Peter to be taken in by
this staging; in fact, if he really believed that she was in the process of
mending her hair-drier herself, he would probably not interfere. She does
expect him to be clever enough to work out that this is a staging intended
to inform him of the fact that she needs some help with her hair-drier.
However, she does not expect him to be clever enough to work out that
she expected him to reason along just these lines. Since she is not really
asking, if Peter fails to help, it will not really count as a refusal either.
This example fits both Grice's original analysis of speaker's meaning
and the reformulations in (27) and (31). Mary does intend Peter to be
informed of her need by recognising her intention to inform him of it. Yet
there is an intuitive reluctance to say that Mary meant that she wanted
Peter's help, or that she was communicating with Peter in the sense we are
trying to characterise. This reluctance, which we believe is well-founded,
has to do with the fact that Mary's second-order intention to have her
first-order informative intention recognised is hidden from Peter.
To deal with such counterexamples, Strawson (1964a), who first drew
attention to the problem, argued that the Gricean analysis must be
enriched: true communication must be characterised as wholly overt. The
question then is how to modify the analysis of inferential communication
to include this requirement of overtness; in other words, how should the
intuitive and rather vague notion of overtness be made more precise?
Answers to this question have been highly technical.
Strawson's own solution was to add to the analysis of speaker's meaning
a third-order intention to have the second-order intention recognised by
the audience; a meta-communicative intention, so to speak, was added to
the informative and communicative intentions. As Straws on envisaged,
and as Schiffer (1972: chapter 2) showed, this is not enough: examples can
Problems of explanation
31
be constructed where the third-order meta-communicative intention is
present but hidden from the audience, and the resulting interaction lacks
the required overtness. Adding a fourth-order meta-meta-communicative
intention that the third-order meta-communicative intention should itself
be recognised by the audience may not be enough either: in principle, for
any nth-order intention of this type, you need an n+ lth-order intention
to the effect that the nth-order intention be recognised. In other words
you need an infinity of such intentions to explicate the intuitive notion of
overtness along those lines.
There are ways of making logical sense of an infinity of intentions, and
of analysing speaker's meaning or communication in terms of such an
infinityY But the results have little psychological plausibility. Fr?m the
psychological point of view, intentions are mental representations capable
of being realised in the form of actions. No psychologist would want to
analyse an utterance as the realisation of an infinity of intentions so
understood. 20
The intuitive idea that communicative intentions must be overt can be
worked out in another way, using the notion of mutual knowledge. This
solution, proposed by Schiffer (1972), essentially involves the assumption
that a true communicative intention is not just an intention to inform the
audience of the communicator's informative intention, but an intention to
make the informative intention mutually known to the communicator and
the audience. By this criterion, the counterexample of Mary trying to get
Peter to repair her hair-drier without openly asking him is not a case of
true communication. Although Mary wants Peter to recognise her
informative intention, she does not want this informative intention to
become mutually known to both of them. More complex examples built
on the same pattern would similarly be ruled out by this mutualknowledge requirement. 21
We have already argued (in section 3) that the appeal to 'mutual
knowledge' lacks psychological plausibility. Hence to rely on it in
explicating the notion of overtness is to turn one's back on psychology
once more. Thus, all the solutions to the overtness problem proposed so
far replace vagueness by one inadequate formalism or another. What we
believe is a satisfactory solution will be proposed in section 8 and
developed in section 12. In the meantime, we turn to further problems
with Grice's analysis, problems this time not of definition but of
explanation.
7 Problems of explanation: Grice's theory of conversation

The Gricean analysis of communication has been discussed almost
exclusively by philosophers,22 whose main concern has been to define the
32
Communication
terms 'meaning' or 'communication'. From our current, more psycholo_

gical point of view, defining communication is not a primary concern. For
one thing, communication does not necessarily involve a distinct and
homogeneous set of empirical phenomena. Our aim is to identify
underlying mechanisms, rooted in human psychology, which explain how
humans communicate with one another. A psychologically well-founded
definition and typology of communication, if possible at all, should
follow from a theoretical account of these underlying mechanisms. We see
Grice's analysis as a possible basis for such a theoretical account. From this
perspective, the main defect of Grice's analysis i~ not that it defines communication too vaguely, but that it explains communication too poorly.
The code model has the merit of explaining how communication could
in principle be achieved. It fails not on the explanatory but on the
descriptive side: humans do not communicate by encoding and decoding
thoughts. The inferential model, despite the technical problems discussed
in the last section, provides a description of human communication which
rings true. By itself, however, it explains very little. The temptation to
return to the code model will remain powerful as long as the inferential
model is not developed into a plausible explanatory account of communication. However, the basis for such an account is suggested by
another work of Grice's, his William James Lectures, in which he puts
forward the view that communication is governed by a 'co-operative
principle' and 'maxims of conversation'.23
According to the inferential model, communication is achieved by the
audience recognising the communicator's informative intention. However, it is not enough to point out, as we have done, that recognising
intentions is a normal feature of human cognition. The recognition of
informative intentions presents problems which the recognition of other
human intentions does not.
How does one recognise another individual's intentions? One observes
his behaviour; using one's knowledge of people in general and of the
individual in particular, one infers which of the effects of this behaviour he
could have both predicted and desired; one then assumes that these
predictable and desirable effects were also intended. In other words, one
infers the intention behind the behaviour from its independently observed
or inferred effects. This pattern of inference is generally not available to an
audience trying to recognise a communicator's informative intention. As
we have seen, the informative effects of communication are normally
achieved, if at all, via recognition of the informative intention. Hence, it
seems, the audience cannot first observe or infer these effects, and then use
them to infer the informative intention.
However, the problem is not that it is hard to come up with hypotheses
33
bout what the communicator might have intended to convey: it is that
:00 many hypothese.s a~e.possible. Even.a lingu~stic utterance ~s generally
full of semantic ambIgUItieS and referential ambIvalences, and IS open to a

wide range of figurative interpretations. For non-coded behaviour there
is, by definition, no predetermined ran~e of information i.t might be used
to communicate. The problem, then, IS to choose the nght hypothesis
from an indefinite range of possible hypotheses. How can this be done?
First, it is easy enough to infer that a certain piece of behaviour is
communicative. Communicative behaviour has at least one characteristic
effect which is achieved before the communicator's informative intention
is recognised: it overtly claims the audience's attention.
Grice's fundamental idea in his William James Lectures is that once a
certain piece of behaviour is identified as communicative, it is reasonable
to assume that the communicator is trying to meet certain general
standards. From knowledge of these general standards, observation of the
communicator's behaviour, and the context, it should be possible to infer
the communicator's specific informative intention. Grice, talking only of
verbal communication, argues,
Our talk exchanges ... are characteristically, to some degree at least,
cooperative efforts; and each participant recognizes in them, to some
extent, a common purpose or set of purposes, or at least a mutually
accepted direction .... at each stage, some possible conversational
moves would be excluded as conversationally unsuitable. We might
then formulate a rough general principle which participants will be
expected (ceteris paribus) to observe, namely: Makeyour conversational contribution such as is required, at the stage at which it occurs,
by the accepted purpose or direction of the talk exchange in which
you are engaged. (Grice 1975: 45)
This Grice calls the co-operative principle. He then develops it into nine
maxims classified into four categories:
Maxims of quantity
1 Make your contribution as informative as is required (for the current
purposes of the exchange).
2 Do not make your contribution more informative than is required.
Maxims of quality
Supermaxim: Try to make your contribution one that is true.
1 Do not say what you believe to be false.
2 Do not say that for which you lack adequate evidence.
Maxim of relation
Be relevant.
34
Communication
Maxims of manner
Supermaxim: Be perspicuous.
1 Avoid obscurity of expression.
2 Avoid ambiguity.
3 Be brief (avoid unnecessary prolixity).
4 Be orderly.
This account of the general standards governing verbal communication
makes it possible to explain how the utterance of a sentence, which
provides only an incomplete and ambiguous representation of a thought,
can nevertheless express a complete and unaml?iguous thought. 24 Of the
various thoughts which the sentence uttered could be taken to represent,
the hearer can eliminate any that are incompatible with the assumption
that the speaker is obeying the co-operative principle and maxims. If only
one thought is left, then the hearer can infer that it is this thought that the
speaker is trying to communicate. Thus, to communicate efficiently, all
the speaker has to do is utter a sentence only one interpretation of which is
compatible with the assumption that she is obeying the co-operative
principle and maxims.
Recall, for instance, our example (16)-(18):
(16) Jones has bought the Times.
(17) Jones has bought a copy of the Times.
(18) Jones has bought the press enterprise which publishes the Times.
There might be situations where only interpretation (17) of the utterance

in (16) would be compatible with the assumption that the speaker does not
say what she believes to be false (first maxim of quality). There might be
situations where only interpretation (18) would be compatible with the
assumption that the speaker is being relevant (maxim of relation). In those
situations, the intended interpretation of (16) can easily be inferred. Hence
the maxims and the inferences they give rise to make it possible to convey
an unambiguous thought by uttering an ambiguous sentence.
Grice's approach to verbal communication also makes it possible to
explain how utterances can convey not just explicit but also implicit
thoughts. Consider dialogue (32):
(32) Peter: Do you want some coffee?

Mary: Coffee would keep me awake.
Suppose that Peter is aware of (33). Then from the assumption explicitly
expressed by Mary's answer, together with assumption (33), he could
infer conclusion (34):
(33) Mary does not want to stay awak~.
(34) Mary does not want any coffee.
35
In just the same way, if Peter is aware of (35), he could infer conclusion
(36):
(35) Mary's eyes remain open when she is awake.
(36) Coffee would cause Mary's eyes to remain open.
Now in ordinary circumstances, Mary would have wanted to communicate (34) but not (36), although both are inferable in the same way from the
thought she has explicitly expressed. This is easily explained on the
assumption that Mary obeys Grice's maxims. The explicit content of her
utterance does not directly answer Peter's question; it is therefore not
relevant as it stands. If Mary has obeyed the maxim 'be relevant', it must be
assumed that she intended to give Peter an answer. Since he can obtain just
the expected answer by inferring (34) from what she said, she must have
intended him to draw precisely this conclusion. There is no parallel reason
to think that she intended Peter to infer (36). Hence, just as the Gricean
maxims help the hearer choose, from among the senses of an ambiguous
sentence, the one which was intended by the speaker, so they help him
choose, from among the implications of the explicit content of an
utterance, the ones which are implicitly conveyed.
Suppose now that the exchange in (32) takes place in the same
circumstances as before, except that Peter has no particu'iar reason
beforehand to assume that Mary does not want to stay awake. Without
this assumption, no answer to his question is derivable from Mary's
utterance, and the relevance of this utterance is not immediately apparent.
One of Grice's main contributions to pragmatics was to show how, in the
event of such an apparent violation of the co-operative principle and
maxims, hearers are expected to make any additional assumptions needed
to dispose of the violation. Here Peter might first adopt (33) as a specific
assumption jointly suggested by the utterance, his knowledge of Mary,
and the general assumption that Mary is trying to be relevant. He might
then infer, as in the previous example, that she does not want any coffee.
To eliminate the apparent violation of the maxims, Peter would have to
assume that Mary had intended him to reason just as he did: that is, that
she was intending to convey implicitly both assumption (33) and
conclusion (34).
Grice calls additional assumptions and conclusions such as (33) and
(34), supplied to preserve the application of the co-operative principle and
maxims, implicatures. Like his ideas on meaning, Grice's ideas on
implicature can be seen as an attempt to build on a commonsense view of
verbal communication by making it more explicit and exploring its
implications. In his William James Lectures, Grice took one crucial step
away from this commonsense view towards theoretical sophistication;
36
Communication
but of course one step is not enough. Grice's account retains much of the
vagueness of the commonsense view. Essential concepts mentioned in the
maxims are left entirely undefined. This is true of relevance, for instance:
hence appeals to the 'maxim of relation' are no more than dressed-up
appeals to intuition. Thus, everybody would agree that, in ordinary
circumstances, adding (33) and (34) to the interpretation of Mary's answer
in (32) makes it relevant, whereas adding (35) and (36) does not. However,
this fact has itself to be explained before it can be used in a genuine
explanation of how Mary's answer is understood.
Grice's view of implicature raises even more b~sic questions. What is
the rationale behind the co-operative principle and maxims? Are there just
the nine maxims Grice mentioned, or might others be needed, as he
suggested himself? It might be tempting to add a maxim every time a
regularity has to be accounted for. 25 However, this would be entirely ad
hoc. What criteria, then, do individual maxims have to meet? Could the
number of maxims be not expanded but reduced ?26
How are the maxims to be used in inference? Grice himself seems to
think that the hearer uses the assumption that the speaker has observed the
maxims as a premise in inference. Others have tried to reinterpret the
maxims as 'conversational postulates' (Gordon and Lakoff 1975), or even
as code-like rules which take semantic representations of sentences and
descriptions of context as input, and yield pragmatic representations of
utterances as output (Gazdar 1979). The flavour of such proposals can be
seen from the following remarks:
The tactic adopted here is to examine some of the data that would, or
should be, covered by Grice's quantity maxim and then propose a
relatively simple formal solution to the problem of describing the
behaviour of that data. This solution may be seen as a special case of
Grice's quantity maxim, or as an alternative to it, or as merely a
conventional rule for assigning one class of conversational meanings
to one class of utterance. (Gazdar 1979: 49)
The pragmatic phenomena amenable to this sort of treatment are rather
limited: they essentially arise when the utterance of a certain sentence is so
regularly correlated with a certain pragmatic interpretation that it makes
sense to set up a rule linking the one to the other. For example, the
~t~erance of (37) regularly suggests (38), the main exception being when
It IS already assumed that (38) is, or might be, false:
(37) Some of the arguments are convincing.
(38) Not all of the arguments are convincing.
The proposal is to deal with this by setting up a general rule associating
37
(37) wit~ t~e pragmatic interI?r~tation (38), and eff:ctively. blocking its
application III contexts where It IS assumed that (38) IS, or mIght be, false
(Gazdar 1979: 55-9). However, in most cases of implicature, as for
instance in example (32)-(34), the context does much more than filter out
inappropriate interpretations: it provides premises without which the
implicature cannot be inferred at all. The translation of Grice's maxims
into code-like rules would thus reduce them to dealing with a narrow set
of interesting but quite untypical examples of implicature.
What, then, are the forms of inference involved in the J:.1ormal operation
of the maxims? If, as seems plausible, non-demonstrative (i.e. nondeductive) inference is involved, how does it operate? Without pursuing
these questions in any depth, most pragmatists have adopted one form or
another of the Gricean approach to implicatures, and are otherwise
content to explain the explicit core of verbal communication in terms of
the code model. The results are as can be expected. Although based on an
insight which seems quite correct, and although somewhat more explicit
and systematic than the intuitive reconstructions supplied by unsophisticated speakers, the analyses of implicature which have been proposed by
pragmatists have shared with these intuitive reconstructions the defect of
being almost entirely ex post facto.
.
Given that an utterance in context was found to carry particular
implicatures, what both the hearer and the pragmatic theorist can do, the
latter in a slightly more sophisticated way, is to show how in very intuitive
terms there was an argument based on the context, the utterance and
general expectations about the behaviour of speakers, that would justify
the particular interpretation chosen. What they fail to show is that on the
same basis, an equally convincing justification could not have been given
for some other interpretation that was not in fact chosen. There may be a
whole variety of interpretations that would meet whatever standards of
truthfulness, informativeness, relevance and clarity have been proposed or
envisaged so far. The theory needs improving at a fundamental level
before it can be fruitfully applied to particular cases.
In his William James Lectures, Grice put forward an idea of fundamental importance: that the very act of communicating creates
expectations which it then exploits. Grice himself first applied this idea
and its elaboration in terms of the maxims to a rather limited problem of
linguistic philosophy: do logical connectives ('and', 'or', 'if ... then') have
the same meaning in natural languages as they do in logic? He argued that
the richer meaning these connectives seem to have in natural languages can
be explained in terms not of word meaning but of implicature. He then
suggested that this approach could have wider applications: that the task
of linguistic semantics could be considerably simplified by treating a large
array of problems in terms of implicatures. And indeed, the study of
38
Communication
implicature along Gricean lines has become a major concern of

pragmatics. 27 We believe that the basic idea of Grice's William James
Lectures has even wider implications: it offers a way of developing the
analysis of inferential communication, suggested by Grice himself in
'Meaning' (1957), into an explanatory model. To achieve this, however,
we must leave aside the various elaborations of Grice's original hunches
and the sophisticated, though empirically rather empty debates they have
given rise to. What is needed is an attempt to rethink, in psychologically
realistic terms, such basic questions as: What form of shared information
is available to humans? How is shared infornption exploited in communication? What is relevance and how is it achieved? What role does the
search for relevance play in communication? It is to these questions that
we now turn.
8 Cognitive environments and mutual manifestness

We have argued that mutual knowledge is a philosopher's construct with
no close counterpart in reality. This is not to deny that humans do, in some
sense, share information. In the first place, the communication process
itself gives rise to shared information; in the second place, some sharing of
information is necessary if communication is to be achieved. Any account
of human communication must thus incorporate some notion of shared
information. In this section, we want to go beyond both the empirically
inadequate notion of 'mutual knowledge' and the conceptually vague
notion of 'shared information' . We will discuss in what sense humans
share information, and to what extent they share information about the
information they share.
All humans live in the same physical world. We are all engaged in a
lifetime's enterprise of deriving information from this common environment and constructing the best possible mental representation of it. We do
not all construct the same representation, because of differences in our
narrower physical environments on the one hand, and in our cognitive
abilities on the other. Perceptual abilities vary in effectiveness from one
individual to another. Inferential abilities also vary, and not just in
effectiveness. People speak different languages, they have mastered
different concepts; as a result, they can construct different representations
and make different inferences. They have different memories, too,
different theories that they bring to bear on their experience in different
ways. Hence, even if they all shared the same narrow physical environment, what we propose to call their cognitive environments would still
differ.
To introduce the notion of a cognitive environment, let us consider a
39
parallel case. O?~ hu~an cogn~tive ability i~ sight. With respec~ to sight,
each individual IS m a vIsual enVIronment whIch can be charactensed as the
set of all phenomena visible to him. What is visible to him is a function
both of his physical environment and of his visual abilities.
In studying communication, we are interested in conceptual cognitive
abilities. We want to suggest that what visible phenomena are for visual
cognition, manifest facts are for conceptual cognition. Let us define:
(39) A fact is manifest to an individual.at a .given time if and only ~f he. is
capable at that time of representmg It mentally and acceptmg Its
representation as true or probably true.
(40) A cognitive environment of an individual is a set of facts) that are
manifest to him.
To be manifest, then, is to be perceptible or inferable. 28 An individual's

total cognitive environment is the set of all the facts that he can perceive or
infer: all the facts that are manifest to him. An individual's total cognitive
environment is a function of his physical environment and his cognitive
abilities. It consists of not only all the facts that he is aware of, but also all
the facts that he is capable of becoming aware of, in his physical
environment. The individual's actual awareness of facts, i.e. the knowledge that he has acquired, of course contributes to his ability to become
aware of further facts. Memorised information is a component of
cognitive abilities.
We want to elaborate the notion of what is manifest in two ways: first,
we want to extend it from facts to all assumptions; and second, we want to
distinguish degrees of manifestness. Our point of view here is cognitive
rather than epistemological. From a cognitive point of view, mistaken
assumptions can be indistinguishable from genuine factual knowledge,
just as optical illusions can be indistinguishable from true sight. Just as
illusions are 'visible', so any assumption, whether true or false, may be
manifest to an individual. An assumption, then, is manifest in a cognitive
environment if the environment provides sufficient evidence for its
adoption, and as we all know, mistaken assumptions are sometimes very
well evidenced.
Anything that can be seen at all is visible, but some things are much
more visible than others. Similarly, we have defined 'manifest' so that any
assumption that an individual is capable of constructing and accepting as
true or probably true is manifest to him. We also want to say that manifest
assumptions which are more likely to be entertained are more manifest.
Which assumptions are more manifest to an individual during a given
period or at a given moment is again a function of his physical
environment on the one hand and his cognitive abilities on the other.
40
Communication
Human cognitive organisation makes certain types of phenomena (i.e.

perceptible objects or events) particularly salient. For instance, the noise
of an explosion or a doorbell ringing is highly salient, a background buzz
or a ticking clock much less so. When a phenomenon is noticed, some
assumptions about it are standardly more accessible than others. In an
environment where the doorbell has just rung, it will normally be strongly
manifest that there is someone at the door, less strongly so that whoever is
at the door is tall enough to reach the bell, and less strongly still that the
bell has not been stolen. The most strongly manifest assumption of all is
the assumption that the doorbell has just rung, fhe evidence for which is
both salient and conclusive. We will have more to say, in chapter 3, about
the factors which make some assumptions more manifest than others in a
given situation. For the moment it is the fact rather than the explanation
that matters.
Our notion of what is manifest to an individual is clearly weaker than
the notion of what is actually known or assumed. A fact can be manifest
without being known; all the individual's actual assumptions are manifest
to him, but many more assumptions which he has not actually made are
manifest to him too. This is so however weakly the terms 'knowledge' and
'assumption' are construed. In a strong sense, to know some fact involves
having a mental representation of it. In a weaker sense, to say that an
individual knows some fact is not necessarily to imply that he has ever
entertained a mental representation of it. For instance, before reading this
sentence you all knew, in that weak sense, that N oam Chomsky never had
breakfast with Julius Caesar, although until now the thought of it had
never crossed your mind. It is generally accepted that people have not only
the knowledge that they actually entertain, but also the knowledge that
they are capable of deducing from the knowledge that they entertain.
However, something can be manifest without being known, even in this
virtual way, if only because something can be manifest and false, whereas
nothing can be known and false.
Can something be manifest without being actually assumed? The
answer must again be yes. Assumptions are unlike knowledge in that they
need not be true. As with knowledge, people can be said to assume, in a
weak sense, what they are capable of deducing from what they assume.
However, people do not assume, in any sense, what they are merely
capable of inferring non-demonstratively - that is, by some creative
process of hypothesis formation and confirmation - from what they
assume. Although it presumably followed non-demonstratively from
what you knew and assumed before you read this sentence that Ronald
Reagan and Noam Chomsky never played billiards together, this was not,
until now, an assumption of yours: it was only an assumption that was
manifest to you. Moreover, something can be manifest merely by being
41
perceptible, and without being inferable at all from previously held

knowledge and assumptions. A car is audibly passing in the street. You
have not yet paid any attention to it, so you have no knowledge of it, no
assumptions about it, even in the weakest sense of 'knowledge' and
'assumption'. But the fact that a car is passing in the street is manifest to
f
. weak er t h an 'known' or
you.
We Wl11 now sh ow t h at b ecause 'mam
est' IS
'assumed', a notion of mutual manifestness can be developed which does
not suffer from the same psychological implausibility as 'mutual knowledge' or 'mutual assumptions'.
To the extent that two organisms have the same visual abilities and the
same physical environment, the same phenomena are visible to them and
they can be said to share a visual environment. Since visual abilities and
physical environments are never exactly identical, organisms never share
their total visual environments. Moreover, two organisms which share a
visual environment need not actually see the same phenomena; they are
merely capable of doing so.
Similarly, the same facts and assumptions may be manifest in the
cognitive environments of two different people. In that case, these
cognitive environments intersect, and their intersection is a cognitive
environment that these two people share. The total shared cognitive
environment of two people is the intersection of their two total cognitive
environments: i.e. the set of all facts that are manifest to them both.
Clearly, if people share cognitive environments, it is because they share
physical environments and have similar cognitive abilities. Since physical
environments are never strictly identical, and since cognitive abilities are
affected by previously memorised information and thus differ in many
respects from one person to another, people never share their total
cognitive environments. Moreover, to say that two people share a
cognitive environment does not imply that they make the same assumptions: merely that they are capable of doing so.
One thing that can be manifest in a given cognitive environment is a
characterisation of the people who have access to it. For instance, every
Freemason has access to a number of secret assumptions which include the
assumption that all Freemasons have access to these same secret assumptions. In other words, all Freemasons share a cognitive environment
which contains the assumption that all Freemasons share this environment. To take another example, Peter and Mary are talking to each other
in the same room: they share a cognitive environment which consists of all
the facts made manifest to them by their presence in this room. One of
these facts is the fact that they share this environment.
Any shared cognitive environment in which it is manifest which people
share it is what we will call a mutual cognitive environment. In a mutual
Communication
cognitive environment, for every manifest assumption, the fact that it is
manifest to the people who share this environment is itself manifest. In
other words, in a mutual cognitive environment, every manifest assumption is what we will call mutually manifest.
Consider, for example, a cognitive environment E shared by Peter and
Mary, in which (41) and (42) are manifest:
42
(41) Peter and Mary share cognitive environment E.

(42) The phone is ringing.
In this environment, (43)-(45) and indefinitely many assumptions built on
the same pattern are also mamfest:
(43) It is manifest to Peter and to Mary that the phone is ringing.

(44) It is manifest to Peter and to Mary that it is manifest to Peter and to
Mary that the phone is ringing.
(45) It is manifest to Peter and to Mary that it is manifest to Peter and to
Mary that it is manifest to Peter and to Mary that the phone is ringing.
The more complex assumptions of type (43)-(45) get, the less likely they
are actually to be made. However, in such a series, assumption n does not
have to be actually made by the individuals it mentions for assumption
n+ 1 to be true. There is therefore no cut-off point beyond which these
assumptions are likely to be false rather than true; they remain manifest
throughout, even though their degree of manifestness tends asymptotically toward zero. (41 )-( 45) and all the assumptions in E are not only
manifest to Peter and Mary; they are mutually manifest.
The notion of a mutually manifest assumption is clearly weaker than
that of a mutual assumption (and a fortiori than that of mutual
knowledge). Consider assumptions (46)-(48) and all the further assumptions that can be built on the same pattern:
(46) Peter and Mary assume that the phone is ringing.
(47) Peter and Mary assume that Peter and Mary assume that the phone is
rIngmg.
(48) Peter and Mary assume that Peter and Mary assume that Peter and
Mary assume that the phone is ringing.
As before, the more complex assumptions of type (46)-(48) get, the less
likely they are actually to be made. In this case, however, assumption n
does have to be made by Peter and Mary for assumption n + 1 to be true.
Moreover, there is sure to be some point - quite soon actually - at which
Mary does not assume that Peter assumes that she assumes that he assumes,
etc. At this point and beyond, all the assumptions in this series are false, and
mutuality of assumptions is not achieved. Another way of seeing that mutuality of assumptions is stronger than mutual manifestness is to notice that
43
(43) may be true when (46) is not, (44) may be true when (47) is not, (45)
may be true when (48) is not, and so on, while the converse is not possible.
Mutual manifestness is not merely weaker than mutual knowledge or
mutual assumption; it is weaker in just the right way. On the one hand, it
is not open to the same psychological objections, since the claim that an
assumption is mutually manifest is a claim about cognitive environments
rather than mental states or processes. On the other hand, as we will show
in section 12, the notion of mutual manifestness is strong enough to give a
precise and interesting content to the notion of overtness discussed in
section 6. However, by rejecting the notion of mutual knowledge and
adopting the weaker notion ?f ~utual manifestness, we 1ep:ive ourselves
of a certain type of explanatIOn III the study of commUnICatIOn.
Communication requires some degree of co-ordination between communicator and audience on the choice of a code and a context. The notion
of mutual knowledge is used to explain how this co-ordination can be
achieved: given enough mutual knowledge, communicator and audience
can make symmetrical choices of code and context. A realistic notion of
mutual manifestness, on the other hand, is not strong enough to explain
such symmetrical co-ordination. However, before concluding that
mutual manifestness is too weak after all, ask yourself what are the
grounds for assuming that responsibility for co-ordination is equally
shared between communicator and audience, and that both must worry,
symmetrically, about what the other is thinking. Asymmetrical coordination is often easier to achieve, and communication is an asymmetrical process anyhow.
Consider what would happen in ballroom dancing if the responsibility
for choosing steps was left equally to both partners (and how little help the
mutual-knowledge framework would be for solving the resulting coordination problems in real time). Co-ordination problems are avoided,
or considerably reduced, in dancing, by leaving the responsibility to one
partner who leads, while the other has merely to follow. We assume that
the same goes for communication. It is left to the communicator to make
correct assumptions about the codes and contextual information that the
audience will have accessible and be likely to use in the comprehension
process. The responsibility for avoiding misunderstandings also lies with
the speaker, so that all the hearer has to do is go ahead and use whatever
code and contextual information come most easily to hand.
Suppose Mary and Peter are looking at a landscape where she has
noticed a distant church. She says to him,
(49) I've been inside that church.
She does not stop to ask herself whether he has noticed the building, and
whether he assumes she has noticed, and assumes she has noticed he has
Communication
noticed, and so on, or whether he has assumed it is a church, and assumes
she assumes it is, and so on. All she needs is reasonable confidence that he
will be able to identify the building as a church when required to: in other
words, that a certain assumption will be manifest in his cognitive
environment at the right time. He need not have accessed this assumption
before she spoke. In fact, until she spoke he might have thought the
building was a castle: it might be only on the strength of her utterance that
it becomes manifest to him that the building is a church.
Inspired by the landscape, Mary says,
44
(50) It's the sort of scene that would have made Marianne Dashwood
swoon.
This is an allusion to Jane Austen's Sense and Sensibility, a book she
knows Peter has read. She does not stop to think whether he knows she
has read it too and knows she knows he has read it, and so on. Nor is she
unaware of the fact that they may well have reacted to the book in different
ways and remember it differently. Her remark is based on assumptions
that she does not mention and that he need never have made himself before
she spoke. What she expects, rightly, is that her utterance will act as a
prompt, making him recall parts of the book that he had previously
forgotten, and construct the assumptions needed to understand the
allusion.
In both these examples Mary makes assumptions about what assumptions are, or will be, manifest to Peter. Peter trusts that the assumptions he
spontaneously makes about the church and about Sense and Sensibility,
which help him understand Mary's utterances, are those she expected him
to make. To communicate successfully, Mary had to have some knowledge of Peter's cognitive environment. As a result of their successful
communication, their mutual cognitive environment is enlarged. Note
that symmetrical co-ordination and mutual knowledge do not enter into
the picture at all.
The most fundamental reason for adopting the mutual-knowledge
framework, as for adopting the code model, is the desire to show how
successful communication can be guaranteed, how there is some failsafe
algorithm by which the hearer can reconstruct the speaker's exact
meaning. Within this framework the fact that communication often fails is
explained in one of two ways: either the code mechanism has been
imperfectly implemented, or there has been some disruption due to
'noise'. A noiseless, well-implemented code mechanism should guarantee
perfect communication.
In rejecting the mutual-knowledge framework, we abandon the
possibility of using a failsafe algorithm as a model of human communication. But since it is obvious that the communication process takes place at
45
a risk, why assume that it is governed by a failsafe procedure? Moreover, if

there is one conclusion to be drawn from work on artificial intelligence, it
is that most cognitive processes are so complex that they must be modelled
in terms of heuristics rather than failsafe algorithms. We assume, then, that
communication is governed by a less-than-perfect heuristic. On this
approach, failures in communication are to be expected: what is mysterious and requires explanation is not failure but success.
As we have seen, the notion of mutual manifestness is not strong
enough to salvage the code theory of communication. But then, this was
never one of our aims. Instead of taking the code theory for granted and
concluding that mutual knowledge must therefore exist, we prefer to look
at what kind of assumptions people are actually in a position to make
about each other's assumptions, and then see what this implies for an
account of communication.
Sometimes, we have direct evidence about other people's assumptions:
for instance, when they tell us what they assume. More generally, because
we manifestly share cognitive environments with other people, we have
direct evidence about what is manifest to them. When a cognitive
environment we share with other people is mutual, we have evidence
about what is mutually manifest to all of us. Note that this evidence can
never be conclusive: the boundaries of cognitive environments cannot be
precisely determined, if only because the threshold between very weakly
manifest assumptions and inaccessible ones is unmarked.
From assumptions about what is manifest to other people, and in
particular about what is strongly manifest to them, we are in a position to
derive further, though necessarily weaker, assumptions about what
assumptions they are actually making. From assumptions about what is
mutually manifest to all of us, we are in a position to derive further, and
weaker, assumptions about the assumptions they attribute to us. And
essentially, this is it. Human beings somehow manage to communicate in
situations where a great deal can be assumed about what is manifest to
others, a lot can be assumed about what is mutually manifest to themselves
and others, but nothing can be assumed to be truly mutually known or
assumed.
The situations which establish a mutual cognitive environment are
essentially those that have been treated as establishing mutual
knowledge. 29 We have argued that assumptions of mutual knowledge are
never truly warranted. Exa~ples (49) and (50) are anecdotal evidence that
they are unnecessary. The detour via mutual knowledge is superfluous:
mutual cognitive environments directly provide all the information
needed for communication and comprehension. a
The notions of cognitive environment and of manifestness, mutual or
otherwise, are psychologically realistic, but by themselves shed little light
46
Communication
on what goes on in human minds. A cognitive environment is merely a set

of assumptions which the individual is capable of mentally representing
and accepting as true. The question then is: which of these assumptions
will the individual actually make? This question is of interest not only to
the psychologist, but also to every ordinary communicator. We will argue
that when you communicate, your intention is to alter the cognitive
environment of your addressees; but of course you expect their actual
thought processes to be affected as a result. In the next section we will
argue that human cognition is relevance-oriented, and that as a result,
someone who knows an individual's cogniti'oie environment can infer
which assumptions he is actually likely to entertain.
9 Relevance and ostension

An individual's cognitive environment is a set of assumptions available to
him. Which particular assumptions is he most likely to construct and
process? There may, of course, be no general answer to this question. We
want to argue that there is. This book is essentially an exploration of the
idea that there is a single property - relevance - which makes information
worth processing for a human being. Chapter 3 will contain a relatively
technical discussion of relevance. In this section, we simply want to
characterise the notion in very general, informal terms, and to make some
suggestions about the role of relevance in communication.
Human beings are efficient information-processing devices. This is
their most obvious asset as a species. But what is efficiency in information
processing?
Efficiency can only be defined with respect to a goal. Some goals, such
as catching a prey, winning a game or solving a problem, are absolute: they
consist in bringing about a particular state of affairs which at any given
moment either exists or does not exist. Other goals, such as multiplying
one's offspring, improving one's backstroke, or understanding oneself,
are relative: they consist in raising the value of some variable, and can thus
only be achieved to a degree. Efficiency with respect to absolute goals is
simply a matter of reaching them with the smallest possible expenditure of
whatever resource (time, money, energy ... ) it takes. Efficiency with
respect to relative goals is a matter of striking a balance between degree of
achievement and expenditure. In the special case where the expenditure is
fixed - say all the time available is going to be spent anyhow - efficiency
consists in achieving the goal to the highest possible degree.
Most discussions of information processing, whether in experimental
psychology or in artificial intelligence, have been concerned with the
realisation of absolute goals. 'Problem solving' has become the paradigm
47
of information processing. The problems considered have a fixed

solution; the goal of the information-processing device is to find this
solution; efficiency consists in finding it at the minimal cost. However,
not all cognitive tasks fit this description; many tasks consist not in
reaching an absolute goal, but in improving on an existing state of affairs.
Hence, cognitive efficiency may have to be characterised differently for
different devices.
Simpler information-processing devices, whether natural, such as a
frog, or artificial, such as an electronic alarm system, process only very
specific information: for example, metabolic changes and fly movements
for frogs, noises and other vibrations for alarm systems. Their
information-processing activity consists in monitoring changes in the
values of a few variables. They could be informally described as engaged in
answering a few set questions: 'Is there a fly-like object within reach?', 'Is
there a large body moving in the room?' More complex informationprocessing devices, by contrast, can define and monitor new variables or
formulate and answer new questions.
For the simpler devices, efficiency consists in answering their set
questions at the minimal processing cost. Efficiency cannot be so easily
defined for more complex devices such as human beings. For such devices,
efficient information processing may involve formulating and trying to
answer new questions despite the extra processing costs incurred.
Formulating and answering specific questions must then be seen as
subservient to a more general and abstract goal. It is in relation to this
general goal that the efficiency of complex information-processing devices
must be characterised.
On the general goal of human cognition, we have nothing better to offer
than rather trivial speculative remarks. However, these remarks have
important and non-trivial consequences. It seems that human cognition is
aimed at improving the individual's knowledge of the world. This means
adding more information, information that is more accurate, more easily
retrievable, and more developed in areas of greater concern to the
individual. Information processing is a permanent life-long task. An
individual's overall resources for information processing are, if not quite
fixed, at least not very flexible. Thus, long-term cognitive efficiency
consists in improving one's knowledge of the world as much as possible
given the available resources.
What, then, is short-term cognitive efficiency - efficiency, say, in the
way your mind spends the next few seconds or milliseconds? This is a
more concrete question, and one that is harder to answer. At every
moment, many different cognitive tasks could be performed, and this for
two reasons: first, human sensory abilities monitor much more information than central conceptual abilities can process; and second, central
48
Communication
abilities always have plenty of unfinished business. The key problem for
efficient short-term information processing is thus to achieve an optimal
allocation of central processing resources. Resources have to be allocated
to the processing of information which is likely to bring about the greatest
contribution to the mind's general cognitive goals at the smallest
processing cost.
Some information is old: it is already present in the individual's
representation of the world. Unless it is needed for the performance of a
particular cognitive task, and is easier to access from the environment than
from memory, such information is not wortp. processing at all. Other
information is not only new but entirely unconnected with anything in the
individual's representation of the world. It can only be added to this
representation as isolated bits and pieces, and this usually means too much
processing cost for too little benefit. Still other information is new but
connected with old information. When these interconnected new and old
items of information are used together as premises in an inference process,
further new information can be derived: information which could not
have been inferred without this combination of old and new premises.
When the processing of new information gives rise to such a multiplication effect, we call it relevant. The greater the multiplication effect, the
greater the relevance.
Consider an example. Mary and Peter are sitting on a park bench. He
leans back, which alters her view. By leaning back, he modifies her
cognitive environment; he reveals to her certain phenomena, which she
may look at or not, and describe to herself in different ways. Why should
she pay attention to one phenomenon rather than another, or describe it to
herself in one way rather than another? In other words, why should she
mentally process any of the assumptions which have become manifest or
more manifest to her as a result of the change in her environment? Our
answer is that she should process those assumptions that are most relevant
to her at the time.
Imagine, for instance, that as a result of Peter's leaning back she can see,
among other things, three people: an ice-cream vendor who she had
noticed before when she sat down on the bench, an ordinary stroller who
she has never seen before, and her acquaintance William, who is coming
towards them and is a dreadful bore. Many assumptions about each of
these characters are more or less manifest to her. She may already have
considered the implications of the presence of the ice-cream vendor when
she first noticed him; if so, it would be a waste of processing resources to
pay further attention to him now. The presence of the unknown stroller is
new information to her, but little or nothing follows from it; so there
again, what she can perceive and infer about him is not likely to be of much
relevance to her. By contrast, from the fact that William is coming her
49
way, she can draw many conclusions from which many more conclusions
will follow. This, then, is the one truly relevant change in her cognitive
environment; this is the particular phenomenon she should pay attention
to. She should do so, that is, if she is aiming at cognitive efficiency.
Our claim is that all human beings automatically aim at the most
efficient information processing possible. This is so whether they are
conscious of it or not; in fact, the very diverse and shifting conscious
interests of individuals result from the pursuit of this permanent aim in
changing conditions. In other words, an individual's particular cognitive
goal at a given moment is always an instance- of a more general goal:
maximising the relevance of the information processed. We will show that
this is a crucial factor in human interaction.
Among the facts made manifest to Mary by Peter's behaviour is the very
fact that he has behaved in a certain way. Suppose now that she pays
attention to this behaviour, and comes to the conclusion that it must have
been deliberate: perhaps he is leaning back more rigidly than if he were
merely trying to find a more comfortable position. She might then ask
herself why he is doing it. There may be many possible answers; suppose
that the most plausible one she can find is that he is leaning back in order to
attract her attention to some particular phenomenon. Then Peter's
behaviour has made it manifest to Mary that he intends to make some
particular assumptions manifest to her. We will call such behaviour behaviour which makes manifest an intention to make something manifest
- ostensive behaviour or simply ostension. Showing someone something is
a case of ostension. So too, we will argue, is human intentional
communication.
The existence of ostension is beyond doubt. What is puzzling is how it
works. Any perceptible behaviour makes manifest indefinitely many
assumptions. How is the audience of an act of ostension to discover which
of them have been intentionally made manifest? For instance, how is Mary
to discover which of the phenomena which have become manifest to her as
a result of Peter's behaviour are the ones he intended her to pay attention
to?
Information processing involves effort; it will only be undertaken in the
expectation of some reward. There is thus no point in drawing someone's
attention to a phenomenon unless it will seem relevant enough to him to
be worth his attention. By requesting Mary's attention, Peter suggests that
he has reason to think that by paying attention, she will gain some relevant
information. He may, of course, be mistaken, or trying to distract her
attention from relevant information elsewhere, as the maker of an
assertion may be mistaken or lying; but just as an assertion comes with a
tacit guarantee of truth, so ostension comes with a tacit guarantee of
relevance~
50
Communication
This guarantee of relevance makes it possible for Mary to infer which of

the newly manifest assumptions have been intentionally made manifest.
Here is how the inference process might go. First, Mary notices Peter's
behaviour and assumes that it is ostensive: i.e. that it is intended to attract
her attention to some phenomenon. If she has enough confidence in his
guarantee of relevance, she will infer that some of the information which
his behaviour has made manifest to her is indeed relevant to her. She then
pays attention to the area that has become visible to her as a result of his
leaning back, and discovers the ice-cream vendor, the stroller, this
dreadful William, and so on. Assumptions about William are the only
newly manifest a ,sumptions relevant enough to be worth her attention.
From this, she can infer that Peter's intention was precisely to draw her
attention to William's arrival. Any other assumption about his ostensive
behaviour is inconsistent with her confidence in the guarantee of relevance
it carries.
Mary has become aware not only that there is someone coming who she
wants to avoid, but also that Peter intended her to become aware of it, and
that he is aware of it too. On the basis of his observable behaviour, she has
discovered some of his thoughts.
Ostensive behaviour provides evidence of one's thoughts. It succeeds in
doing so because it implies a guarantee of relevance. It implies such a
guarantee because humans automatically turn their attention to what
seems most relevant to them. The main thesis of this book is that an act of
ostension carries a guarantee of relevance, and that this fact - which we
will call the principle of relevance - makes manifest the intention behind
the ostension. eWe believe that it is this principle of relevance that is
needed to make the inferential model of communication explanatory.
10 Ostensive-inferential communication
Ostension provides two layers of information to be picked up: first, there
is the information which has been, so to speak, pointed out; second, there
is the information that the first layer of information has been intentionally
pointed out. One can imagine the first layer being recovered without the
second. For example, as a result of Peter's leaning back, Mary might notice
William coming their way, even if she paid no attention to Peter's
intentions. And as for Peter, he might not care much whether Mary
recognises his intention, as long as she notices William.
In general, however, recognising the intention behind the ostension is
necessary for efficient information processing: someone who fails to
recognise this intention may fail to notice relevant information. Let us
51
modify our example slightly and suppose that William is in the distance,
barely visible in a crowd. If Mary pays no attention to the fact that Peter's
behaviour is ostensive, she might well look in the right direction and yet
not notice William. If she pays attention to the ostension, she will be
inclined to take a closer look and find out what information Peter thought
might be relevant to her.
In our modified example, what Peter's ostension mostly does is make
much more manifest some information which would have been manifest
anyhow, though very weakly so. Sometimes, however, part of the basic
information will not be manifest at all unless the intention behind the
ostension is taken into account. Suppose a girl is travelling in a foreign
country. She comes out of the inn wearing light summer clothes,
manifestly intending to take a stroll. An old man sitting on a bench n~earby
looks ostensively up at the sky. When the girl looks up, she sees a few tiny
clouds, which she might have noticed for herself, but which she would
normally have paid no further attention to: given her knowledge - or lack
of knowledge - of the local weather, the presence of these tiny clouds is
not relevant to her. Now, however, the old man is drawing her attention to
the clouds in a manifestly intentional way, thus guaranteeing that there is
some relevant information to be obtained.
The old man's ostensive behaviour opens up for the girl a whole new
strategy of processing. If she accepts his guarantee of relevance, she has to
find out what makes him think that the presence of the clouds would be
relevant to her. Knowing the area and its weather better than she does, he
might have reason to think that the clouds are going to get worse and turn
to rain. Such an assumption is of a very standard sort and would probably
be the first to come to mind. The old man can thus be reasonably confident
that, prompted by his behaviour, she will have no difficulty in deciding
that this is what he believes. If it were not manifest to the old man that it
was going to rain, it would be hard to explain his behaviour at all. The girl
thus has reason to think that in drawing her attention to the clouds, he
intended to make manifest to her that he believed it was going to rain. As a
result of this act of ostension, she now has some information that was not
available to her before: that he thinks it is going to rain, and hence that
there is a genuine risk of rain.
In this example, the state of affairs that the old man drew the girl's
attention to had been partly manifest to her, and partly not. The presence
of the clouds and the fact that clouds may always turn to rain had been
manifest and merely became more so. However, until that moment she
had regarded the fact that the weather was beautiful as strong evidence that
it would not rain. The risk of rain in that particular situation was not
manifest to her at all. In other words, the clouds were already evidence of
52
Communication
oncoming rain, but evidence that was much too weak. The old man made
that evidence much stronger by pointing it out; as his intentions became
manifest, the assumption that it would rain became manifest too.
Sometimes, all the evidence displayed in an act of ostension bears
directly on the agent's intentions. In these cases, only by discovering the
agent's intentions can the audience also discover, indirectly, the basic
information that the agent intended to make manifest. The relation
between the evidence produced and the basic information conveyed is
arbitrary. The same piece of evidence can be used, on different occasions,
to make manifest different assumptions, even qmtually inconsistent
assumptions, as long as it makes manifest the intention behind the
ostension.
Here is an example. Two prisoners, from different tribes with no
common language, are put in a quarry to work back to back breaking
rocks. Suddenly, prisoner A starts putting some distinct rhythm into the.
sound of his hammer - one-two-three, one-two, one-two-three, onetwo - a rhythm that is both arbitrary and noticeable enough to attract the
attention of prisoner B. This arbitrary pattern in the way the rocks are
being broken has no direct relevance for B. However, there are grounds
for thinking that it has been intentionally produced, and B might ask
himself what A's intentions were in producing it. One plausible assumption is that this is a piece of ostensive behaviour: that is, that A intended B
to notice the pattern. This would in turn make manifest A's desire to
interact with B, which in the circumstances would be relevant enough.
Here is a more substantial example. Prisoners A and B are at work in
their quarry, each with a guard at his shoulder, when suddenly the
attention of the guards is distracted. Both prisoners realise that they have a
good chance of escaping, but only if they can co-ordinate their attack and
overpower their guards simultaneously. Here, it is clear what information
would be relevant: each wants to know when the other will start the
attack. Prisoner A suddenly whistles, the prisoners overpower their
guards and escape. Again, there is no need for a pre-existing code
correlating a whistle with the information that now is the moment t9
attack. The information is obvious enough: it is the only information that
A could conceivably have intended to make manifest in the circumstances.
Could not the repetition of such a situation lead to the development of a
code? Imagine that the two prisoners, caught again, find themselves in the
same predicament: again a whistle, again an escape, and again they are
caught. The next time, prisoner B, who has not realised that both guards
are distracted, hears prisoner A whistle: this time, fortunately, B does not
have to infer what the whistle is intended to make manifest: he knows. The
whistle has become a signal associated by an underlying code to the
message 'Let us overpower our guards now!'
53
Inferential theorists might be tempted to see language as a whole as

having developed in this way: to see conventional meanings as growing
out of natural inferences. 3o This is reminiscent of the story of how
Rockefeller became a millionaire. One day, when he was young and very
poor, Ro~kefelle: found a one-cent coin in the street. He ?ought an apple,
polished It, sold It for two cents, bought two apples, polIshed them, sold
them for four cents ... After one month he bought a cart, after two years he
was about to buy a grocery store, when he inherited the fortune of his
millionaire uncle. We will never know how far hominid efforts at
conventionalising inference might have gone towards establishing a
full-fledged human language. The fact is that the development of human
languages was made possible by a specialised biological endowment.
Whatever the origin of the language or code employed, a piece of coded
behaviour may be used ostensively - that is, to provide two layers of
information: a basic layer of information, which may be about anything at
all, and a second layer consisting of the information that the first layer of
information has been intentionally made manifest. When a coded signal,
or any other arbitrary piece of behaviour, is used ostensively, the evidence
displayed bears directly on the individual's intention, and only indirectly
on the basic layer of information that she intends to make manifest. We are
now, of course, dealing with standard cases of Gricean communication.
Is there a dividing line between instances of ostension which one would
be more inclined to describe as 'showing something', and clear cases of
communication where the communicator unquestionably 'means something'? One of Grice's main concerns was to draw such a line: to
distinguish what he called 'natural meaning' - smoke meaning fire, clouds
meaning rain, etc. - from 'non-natural meaning': the word 'fire' meaning
fire, Peter's utterance meaning that it will rain, etc. Essential to this
distinction was the third type of communicator's intention Grice
mentioned in his analysis: a true communicator intends the recognition of
his informative intention to function as at least part of the audience's
reason for fulfilling that intention. In other words, the first, basic, layer of
information must not be entirely recoverable without reference to the
second.
What ~e have tried to show so far in this section is that there are not two
distinct and well-defined classes, but a continuum of cases of ostension
ranging from 'showing', where strong direct evidence for the basic layer of
information is provided, to 'saying that', where all the evidence is indirect.
Even in our very first case of Peter leaning back ostensively to let Mary see
William approaching, it is arguable that some of the basic information is
made manifest indirectly, through Peter's intention being made manifest.
Someone who engages in any kind of ostensive behaviour intentionally
draws some attention to himself and intentionally makes manifest a few
54
Communication
assumptions about himself: for instance, that he is aware of the basic

information involved, and that he is trying to be relevant. Peter's
ostension might make it manifest not just that William is approaching, but
also that Peter expects Mary to be concerned, and that he is concerned too.
Would we want to say, though, that Peter 'meant something' by his
behaviour? Like most English speakers, we would be reluctant to do so;
but this is irrelevant to our pursuit, which is not to analyse ordinary
language usage, but to describe and explain forms of human communication. Our argument at this stage is this: either inferential communication
consists in providing evidence for what the communicator means, in the
sense of 'meaning' which Grice calls 'non-natuhl meaning', and in that
case inferential communication is not a well-defined class of phenomena at
all; or else showing something should be considered a form of inferential
communication, on a par with meaning something by a certain behaviour,
and inferential communication and ostension should be equated.
There are two questions involved here. One is substantive: which
domains of facts are to be described and explained together? Our answer is
that os tension is such a domain, and that inferential communication
narrowly understood (i.e. understood as excluding cases of ostension
where talk of 'meaning' would be awkward) is not. The second question is
terminological (and hence not worth much argument): can the term
'communication' be legitimately applied to all cases of ostension? Our
answer is yes, and from now on we will treat ostensive communication,
inferential communication, and ostensive-inferential communication as
the same thing. Inferential communication and ostension are one and the
same process, but seen from two different points of view: that of the
communicator who is involved in ostension and that of the audience who
is involved in inference.
Ostensive-inferential communication consists in making manifest to an
audience one's intention to make manifest a basic layer of information. It
can therefore be described in terms of an informative and a communicative
intention. In the next two sections, we want to reanalyse the notions of
informative and communicative intention in terms of manifestness and
mutual manifestness, and to sketch in some of the empirical implications
of this reformulation.
11 The informative intention

We began this chapter by pointing out that any account of communication
must answer two questions: first, what is communicated; and second,
how is communication achieved? Up to now, we have considered only the
The informative intention
55
second question. In this section, we return to the first. The generally

accepted answer is that what is communicated is a meaning. The question
then becomes, what is a meaning? And there is no generally accepted
answer any more.
However much they differ, all answers to the what-is-a-meaning
question share the view t~at t~e .paradigm example of meaning is :vha~ is
explicitly expressed by a ImgUIstlc utterance. The verbal commumcatlOn
of an explicit meaning is then taken as the model of communication in
general. This is true of semiotic approaches, which are not only
generalisations of a linguistic model, but are also based on the assumption
that to communicate is always, in Saussure's terms, to trasmit a 'signified'
by use of a 'signifier'. It is true of inferential approaches, which regard all
communicative acts as 'utterances' in an extended sense, used to convey an
'utterer's meaning'.
We believe that the kind of explicit communication that can be achieved
by the use of language is not a typical but a limiting case. Treating
linguistic communication as the model of communication in general has
led to theoretical distortions and misperceptions of the data. The effects of
most forms of human communication, including some of the effects of
verbal communication, are far too vague to be properly analysed along
these lines. Moreover, there is not a dichotomy but a continuum of cases,
from vaguer to more precise effects.
Let us first illustrate this point with two examples of non-verbal
communication. Mary comes home; Peter opens the door. Mary stops at
the door and sniffs ostensively; Peter follows suit and notices that there is
a smell of gas. This fact is highly relevant, and in the absence of contextual
counterevidence or any obvious alternative candidate, Peter will assume
that Mary intended to make it manifest to him that there was a smell of gas.
Here, at least part of what is communicated could be reasonably well
paraphrased by saying that there is a smell of gas; and it could be argued
that this is what Mary means. She could indeed have achieved essentially
the same result by speaking rather than sniffing ostensively.
Contrast this with the following case. Mary and Peter are newly arrived
at the seaside. She opens the window overlooking the sea and sniffs
appreciatively and ostensively. When Peter follows suit, there is no one
particular good thing that comes to his attention: the air smells fresh,
fresher than it did in town, it reminds him of their previous holidays, he
can smell the sea, seaweed, ozone, fish; all sorts of pleasant things come to
mind, and while, because her sniff was appreciative, he is reasonably safe
in assuming that she must have intended him to notice at least some of
them, he is unlikely to be able to pin her intentions down any further. Is
there any reason to assume that her intentions were more specific? Is there
56
Communication
a plausible answer, in the form of an explicit linguistic paraphrase, to the

question, what does she mean? Could she have achieved the same
communicative effect by speaking? Clearly not.
Examples like the one of Mary smelling gas, where it is reasonable to
impute a meaning to the communicator, are the only ones normally
considered in discussions of communication; examples like the one of
Mary at the seaside - clearly communicating, but what? - are generally
ignored. Yet these examples do not belong to distinct classes of
phenomena, and it is easy enough to imagine intermediate cases: say, a
guest sniffing appreciatively and ostensively w~n the stew is brought to
the table, etc.
The distortions and misperceptions introduced by the explicit communication model are also found in the study of verbal communication
itself. Some essential aspects of implicit verbal communication are
overlooked. Pragmatists assume that what is communicated by an
utterance is a speaker's meaning, which in the case of an assertion is a set of
assumptions. One of these assumptions is explicitly expressed; the others
(if any) are implicitly conveyed, or implicated. The only difference
between the explicit content of an utterance and its implicatures is
supposed to be that the explicit content is decoded, while the implicatures
are inferred. Now we all know, as speakers and hearers, that what is
implicitly conveyed by an utterance is generally much vaguer than what is
explicitly expressed, and that when the implicit import of an utterance is
explicitly spelled out, it tends to be distorted by the elimination of this
often intentional vagueness. The distortion is even greater in the case of
metaphor and other figures of speech, whose poetic effects are generally
destroyed by being explicitly spelled out.
In an effort to minimise the distortion, pragmatists have tended to focus
on examples such as (32), where the implicit import is fairly precise, and
to ignore equally ordinary cases of implicit vagueness such as (51):
(32) Peter: Do you want some coffee?

Mary: Coffee would keep me awake.
(51) Peter: What do you intend to do today?
Mary: I have a terrible headache.
In (32), Mary implicates that she doesn't want coffee (or, in some
circumstances, that she does) and that her reason for not wanting it is that
it would keep her awake. Here the implicatures can be spelled out without
distortion. In (51), what does Mary implicate? That she will not do
anything? That she will do as little as possible? That she will do as much as
she can? That she does not yet know what she will do? There is no precise
assumption, apart from the one explicitly expressed, which she can be said
57
to intend Peter to share. Yet there is more to her utterance than its explicit
content: she manifestly intends Peter to draw some conclusions from
what she said, and not just any conclusions. Quite ordinary cases such as
(51) are never discussed in the pragmatic literature.
Pragmatists tend to take for granted that a meaning is a proposition
combined with a propositional attitude, though they may diverge
considerably in the way they present and develop this view. In other
words, they treat the communicator's informative intention as an
intention to induce in an audience certain attitudes to certain propositions.
With assertions, often taken to be the most basic case, the informative
intention is treated as an intention to induce in an audience the belief that a
certain propositIOn is true.
There is a very good reason for anyone concerned with the role of
inference in communication to assume that what is communicated is
propositional: it is relatively easy to say what propositions are, and how
inference might operate over propositions. No one has any clear idea how
inference might operate over non-propositional objects: say, over images,
impressions or emotions. Propositional contents and attitudes thus seem
to provide the only relatively solid ground on which to base a partly or
wholly inferential approach to communication. Too bad if much of what
is communicated does not fit the propositional mould.
At first sight, it might look as if semioticians had a more comprehensive
view. They have an a priori account of how any kind of representation,
propositional or not, might be conveyed: namely, by means of a code.
However, studies by semioticians of what they call 'connotation', i.e. the
vaguer aspect of what is communicated, are highly programmatic and do
not offer the beginnings of a psychologically adequate account of the type
of mental representation involved.}! The semiotic approach is more
comprehensive only by being more superficial.
The only people who have been quite consistently concerned with the
vaguer aspects of communication are the Romantics, from the Schlegel
brothers and Coleridge to 1. A. Richards, and their many acknowledged
or unacknowledged followers, including many semioticians such as
Roman Jakobson in some of his writings, Victor Turner, or Roland
Barthes. However, they have all dealt with vagueness in vague terms, with
metaphors in metaphorical terms, and used the term 'meaning' so broadly
that it becomes quite meaningless.
We see it as a major challenge for any account of human communication
to give a precise description and explanation of its vaguer effects.
Distinguishing meaning from communication, accepting that something can
be communicated without being strictly speaking meant by the communicator or the communicator's behaviour, is a first essential step - a step
away from the traditional approach to communication and most modern
58
Communication
approaches. Once this step is taken, we believe that the framework we

propose, unlike the others we have discussed, can rise to this challenge.
Accounts of communication either are not psychological at all, and
avoid all talk of thoughts, intentions, etc., or else they assume that a
communicator's intention is to induce certain specific thoughts in an
audience. We want to suggest that the communicator's informative
intention is better described as an intention to modify directly not the
thoughts but the cognitive environment of the audience. The actual
cognitive effects of a modification of the cognitive environment are only
partly predictable. Communicators -like human agents in general- form intentions over whose fulfilment they have some control: they can have some
controllable effect on their audience's cognitive environment, much less on
their audience's actual thoughts, and they form their intentions accordingly.
We therefore propose to reformulate the notion of an informative
intention along the following lines. A communicator produces a stimulus
intending thereby
(52) Informative intention: to make manifest or more manifest to the

audience a set of assumptions I.
We take an intention to be a psychological state, and we assume that the
content of the intention must be mentally represented. In particular, the
communicator must have in mind a representation of the set of assumptions I which she intends to make manifest or more manifest to the
audience. However, to have a representation of a set of assumptions it is
not necessary to have a representation of each assumption in the set. Any
individuating description may do.
When the communicator's intention is to make manifest some specific
assumptions, then, of course, her representation of I may be in the form
of a list of assumptions which are members of I. Consider dialogue (53),
for instance:
(53) Passenger: When does the train arrive at Oxford?

Ticket-collector: At 5 :25.
Here the ticket-collector's informative intention is to make manifest to the
passenger the single assumption that the train arrives at 5:25. Examples of
this type, where the communicator wants to communicate one or more
specific assumptions which she actually has in mind, are the only ones
usually considered. Our characterisation (52) of informative intentions
fits these cases quite straightforwardly, but unlike other approaches, is not
limited to them.
Consider, at the other extreme, the vaguest forms of communication.
Here the communicator may have a representation of I in which none of
the assumptions in I is directly listed. For instance, Mary's informative

59
intention when sniffing the seaside air might be that all the assumptions
which became manifest to her when she opened the window and took a
deep breath should, as a result of her ostensive behaviour, become
manifest or more manifest to Peter. She need not intend to communicate
any particular one of these assumptions.
If asked what she wanted to convey, one of the best answers Mary could
give is that she wanted to share an impression with Peter. What is an
impression? Is it a type of mental representation? Can it be reduced to
propositions and propositional attitudes? What we are suggesting is that
an impression might be better described as a noticeable change in one's
cognitive environment, a change resulting from relatively small alterations
in the manifestness of many assumptions, rather than from the fact that a
single assumption or a few new assumptions have all of a sudden become
very manifest. It is quite in line with common sense to think of an
impression as the sort of thing that can be communicated, and yet this
intuition is unexplainable within current theories of communication. In
the model of ostensive-inferential communicadon we are trying to
develop, impressions fall squarely within the domain of things that can be
communicated, and their very vagueness can be precisely described.
In many - perhaps most - cases of human communication, what the
communicator intends to make manifest is partly precise and partly vague.
She may have in mind a characterisation of I based on a representation of
some but not all of the assumptions in 1. For instance, in (51), Mary's
informative intention in saying that she has a headache might be described
as follows: she intends to make manifest to Peter the assumption that she
has a headache and all the further assumptions manifestly required to
make this a relevant answer to Peter's question. Similarly, Mary's
informative intention when sniffing the smell of gas might be to make
manifest to Peter not only the assumption that there is a smell of gas, but
also all the further assumptions that this initial assumption makes
mutually manifest.
Instead of treating an assumption as either communicated or not communicated, we have a set of assumptions which, as a result of communication, become manifest or more manifest to varying degrees. We might
think of communication itself, then, as a matter of degree. When the
communicator makes strongly manifest her informative intention to make
some particular assumption strongly manifest, then that assumption is
strongly communicated. An example would be answering a clear 'Yes'
when asked 'Did you pay the rent?' When the communicator's intention
is to increase simultaneously the manifestness of a wide range of assumptions, so that her intention concerning each of these assumptions is weakly
manifest, then each of them is weakly communicated. An example would
be sniffing ecstatically and ostensively at the fresh seaside air. There is, of
60
Communication
course, a continuum of cases in between. In the case of strong communication, the communicator can have fairly precise expectations about some
of the thoughts that the audience will actually entertain. With weaker
forms of communication, the communicator can merely expect to steer
the thoughts of the audience in a certain direction. Often, in human
interaction, weak communication is found sufficient or even preferable
to the stronger forms.
Non-verbal communication tends to be relatively weak. One of the
advantages of verbal communication is that it gives rise to the strongest
possible form of communication; it enables the pearer to pin down the
speaker's intentions about the explicit content of her utterance to a single,
strongly manifest candidate, with no alternative worth considering at all.
On the other hand, what is implicit in verbal communication is generally
weakly communicated: the hearer can often fulfil part of the speaker's
informative intention by forming any of several roughly similar but not
identical assumptions. Because all communication has been seen as strong
communication, descriptions of non-verbal communication have been
marred by spurious attributions of 'meaning'; in the case of verbal
communication, the difference between explicit content and implicit
import has been seen as a difference not in what gets communicated but
merely in the means by which it is communicated, and the vagueness of
implicatures and non-literal forms of expression has been idealised away.
Our account of informative intentions in terms of manifestness of
assumptions corrects these distortions without introducing either ad hoc
machinery or vagueness of description.
12 The communicative intention

When we introduced the notion of a communicative intention in section 6,
we drew attention to a problem first discussed by Straws on (1964a).
Straws on pointed out that a communicator's intentions must be 'overt' in
a sense which is easy enough to illustrate and grasp intuitively, but hard to
spell out precisely. One type of solution, proposed by Strawson himself, is
to regard an intention as overt when it is backed by a series of further
intentions, each to the effect that the preceding intention in the series
should be recognised. Schiffer (1972) proposed another solution: he
analysed 'overt' as meaning mutually known. We argued that both types
of solution are psychologically implausible.
Our solution, which is closer to Schiffer's than Strawson's, though
without suffering from the defects of either, is to replace the vague 'overt'
by the more precise 'mutually manifest'. We therefore redefine a
communicative intention as follows. To communicate intentionally by
The communicative intention
61
ostension is to produce a certain stimulus with the aim of fulfilling an

informative intention, and intending moreover thereby
(54) Communicative intention: to make it mutually manifest to audience

and communicator that the communicator has this informative
intention.
This takes care of the types of example which Strawson and Schiffer used
to show that, in order to communicate, it is not quite enough to inform an
audience of one's informative intention. For instance, in our example in
section 6, Mary leaves the pieces of her broken hair-drier lying around,
intending thereby to inform Peter that she would like him to mend it. She
wants this informative intention to be manifest to Peter, but at the same
time, she does not want it to be 'overt'. In our terms, she does not want her
informative intention to be mutually manifest. Intuitively, what she does
is not quite communicate. Our redefinition of a communicative intention
accounts for this intuition.
What difference does it make whether an informative intention is
merely manifest to the audience or mutually manifest to audience and
communicator? Should this really be a criterion for distinguishing
communication from other forms of information transmission? Is it more
than a technicality designed to take care of implausible borderline cases
dreamed up by philosophers? Our answer is that there is indeed an
essential difference.
Consider first a more general question: why should someone who has
an informative intention bother to make it known to her audience that she
has this intention? In other words, what are the reasons for engaging in
ostensive communication? Grice discussed only one of these reasons:
sometimes, making one's informative intention known is the best way, or
the only way, of fulfilling it. We have shown that people sometimes
engage in ostensive communication even though the informative intention
could be fulfilled without being made manifest: for example, by providing
direct evidence for the information to be conveyed. However, even in
these cases, ostension helps focus the attention of the audience on the
relevant information, and thus contributes to the fulfilment of the
informative intention. This is still the Gricean reason for engaging in
communication, just slightly extended in scope.
However, we want to argue that there is another major reason for
engaging in ostensive communication, apart from helping to fulfil an
informative intention. Mere informing alters the cognitive environment of
the audience. Communication alters the mutual cognitive environment of
the audience and communicator. Mutual manifestness may be of little
cognitive importance, but it is of crucial social importance. A change in the
mutual cognitive environment of two people is a change in their
Communication
62
possibilities of interaction (and, in particular, m their possibilities of
further communication).
Recall, for instance, the case of Peter leaning back to let Mary see
William coming their way. If, as a result of his behaviour, it becomes
mutually manifest to them that William is coming, that they are in danger
of being bored by his conversation, ete., then they are in a position to act
efficiently: i.e. promptly. All Mary may have to do is say, 'Let's go!'; she
can feel confident that Peter will understand her reasons, and, if he shares
them, will be ready to act without question or delay.
In the case of the broken hair-drier, if Mary had made mutually
manifest her wish that Peter would mend it, one of two things would have
happened. Either he would have mended it, thus granting her wish and
possibly putting her in his debt; or he would have failed to mend it, which
would have amounted to a refusal or rejection. Mary avoids putting
herself in his debt or meeting with a refusal by avoiding any modification
of their mutual cognitive environment. If Peter mends the hair-drier, he is
being kind on his own initiative, and she does not owe him anything. If
Peter decides not to mend the hair-drier, he might reason as follows: she
doesn't know I know she intended to inform me of her wish, so if I ignore
it, she will attribute this to her failure to inform me; she may find me
stupid, but not unkind. As for Mary, she may have intentionally left this
line of reasoning open to Peter. If he does not mend her hair-drier, she will
find him unkind, but not hostile. His failure to grant her wish will not be
in the nature of a rebuff. They will stand in exactly the same social
-relationship to each other as before. This shows how ostensive communication may have social implications that other forms of information
transmission do not.
By making her informative intention mutually manifest, the communicator creates the following situation: it becomes mutually manifest that
the fulfilment of her informative intention is, so to speak, in the hands of
the audience. If the assumptions that she intends to make manifest to the
audience become manifest, then she is successful; if the audience refuses to
accept these assumptions as true or probably true, then she has failed in
her informative intention. Suppose - we will soon see how this may
happen - that the audience's behaviour makes it mutually manifest that the
informative intention is fulfilled. Then the set of assumptions I that the
communicator intended to make manifest to the audience becomes, at
least apparently, mutually manifest. We say 'at least apparently' because,
if the communicator is not sincere and some of the assumptions in I are
not manifest to her, then by our definition of mutual manifestness, these
assumptions cannot be mutually manifest to her and others. 32
A communicator is normally interested in knowing whether or not she
has succeeded in fulfilling her informative intention, and this interest is
The communicative intention
63
mutually manifest to her and her audience. In face-to-face communication, the audience is generally expected to respond to this interest in fairly
conventional ways. Often, for instance, the audience is expected to
communicate its refusal to accept the information communicated, or else
it becomes mutually manifest that the communicator's informative
intention is fulfilled.
Where communication is non-reciprocal, there are various possible
situations to be taken into account. The communicator may be in a
position of such authority over her audience that the success of her
informative intention is mutually manifest in advance. Journalists,
professors, religious or political leaders assume, alas often on good
grounds, that what they communicate automatically becomes mutually
manifest. When the communicator lacks that kind of authority, but still
wants to establish a mutual cognitive environment with her audience, all
she has to do is adapt her informative intentions to her credibility. For
instance, in writing this book we merely intend to make mutually manifest
that we have developed certain hypotheses and have done so on certain
grounds. That is, we take it as mutually manifest that you will accept our
authority on what we actually think. The mutual cognitive environment
thus created is enough for us to go on to communicate further thoughts
which we would otherwise have been unable to communicate. (Of course
we would also like to convince you, but we hope to do this by the force of
our arguments, and not by making you recognise our informative
intentions. )
We began this chapter by asking how human beings communicate with
one another. Our answer is that they use two quite different modes of
communication: coded communication and ostensive-inferential communication. However, the two modes of communication are used in
fundamentally different ways. Whereas ostensive-inferential communication can be used on its own, and sometimes is, coded communication is
only used as a means of strengthening ostensive-inferential communication. This is how language is used in verbal communication, as we will
argue in chapter 4.
Ostensive-inferential communication can be defined as follows:
(55) Ostensive-inferential communication: the communicator produces a

stimulus which makes it mutually manifest to communicator and
audience that the communicator intends, by means of this stimulus,
to make manifest or more manifest to the audience a set of
'1ssumptions {/}.
As this definition stands, it does not exclude the possibility of unintentional communication: that is, a stimulus merely intended to inform might
make mutually manifest the intention to inform, and this, by our
64
Communication
definition, would count as communication. For instance, suppose Mary

yawns, intending to inform Peter that she is tired, and hoping that her
yawn will look natural. She does not do it too well: it is all too obvious that
her yawn is artificial- and her informative intention becomes mutually
manifest. We see no reason for refusing to call this a case of unintended
ostensive communication. It would be easy enough, though, to modify
definition (55) and make intentionality a defining feature of communication.
In any case, most human communication is intentional, and it is
intentional for two good reasons. The first reason is the one suggested by
Grice: by producing direct evidence of one's inf~rmative intention, one
can convey a much wider range of information than can be conveyed by
producing direct evidence for the basic information itself. The second
reason humans have for communicating is to modify and extend the
mutual cognitive environment they share with one another.
What we have offered so far is a good enough description of
ostensive-inferential communication. However, we have not explained
how it works. We have suggested that the explanation is to be sought in a
principle of relevance. To make this principle truly explanatory, we must
first make the notion of relevance much more explicit, and to do this we
must consider how information is mentally represented and inferentially
processed. This, then, is the programme for the next two chapters.
2
Inference
1 Non-demonstrative inference
In the last chapter, we outlined a model of ostensive-inferential communication, looking more closely at the ostensive nature of the communicator's behaviour than at the inferential nature of comprehension. In this
chapter, we will outline a model of the inferential abilities involved in
comprehension. Here, we have already made two broad hypotheses on
which we hope to build. First, we implicitly assumed that the process of
inferential comprehension is non-demonstrative: even under the best of
circumstances, we argued, communication may fail. The addressee can
neither decode nor deduce the communicator's communicative intention.
The best he can do is construct an assumption on the basis of the evidence
provided by the communicator's ostensive behaviour. For such an
assumption, there may be confirmation but no proof.
Second, we explicitly assumed that any conceptually represented
information available to the addressee can be used as a premise in this
inference process. In other words, we assumed that the process of
inferential comprehension is 'global' as opposed to 'local': where a local
process (e.g. deductive reasoning from fixed premises or auditory
perception) is either context-free or sensitive only to contextual information from some set domain, and a global process (e.g. empirical scientific
reasoning) has free access to all conceptual information in memory.
A non-demonstrative inference process with free access to conceptual
memory: this sounds, appropriately enough, like an ordinary central
thought process. A distinction between 'central' processes and 'input',
'perceptual' or 'peripheral' processes is assumed in much of current
cognitive psychology. Roughly speaking, input processes are relatively
specialised decoding processes, whereas central processes are relatively
unspecialised inferential processes. The distinction will be discussed and
illustrated below.
66
Inference
We do maintain that inferential comprehension involves no specialised

mechanisms. In particular, we will argue that the inferential tier of verbal
comprehension involves the application of central, unspecialised inference
processes to the output of specialised, non-inferential linguistic processes.
It seems, then, that our undertaking - and the whole pragmatic enterprise
if our understanding of it is correct - should fall (and we use the word
advisedly) under Fodor's First Law of the Nonexistence of Cognitive
Science, which goes, 'the more global ... a cognitive process is, the less
anybody understands it' (Fodor 1983: 107). a
Fodor points out that while something is known abput the operation of
the perceptual systems, very little is known about the so-called central
thought processes, which integrate information derived from the perceptual systems with information stored in memory, and perform a variety of
inferential tasks. As a typical example of a central thought process, he
takes scientific theorising. The construction and confirmation of a
scientific theory is a global operation in the sense that there is no piece of
evidence, however remote, no hypothesis, however implausible, that
might not turn out to have a bearing on its outcome. It is the global nature
of scientific theorising, Fodor suggests, that makes it so unamenable to
study. To the extent that other central processes share this property, they
are likely to prove equally resistant to investigation:
the reason that there is no serious psychology of central processes is
the same as the reason there is no serious philosophy of scientific
confirmation. Both exemplify the significance of global factors in the
fixation of belief, and nobody begins to understand how such factors
have their effects. (Fodor 1983: 129)
If inferential comprehension is a central thought process, the wish to

construct an adequate theory of ostensive-inferential communication
appears to lead into very deep waters indeed.
We do not entirely share this pessimism. We doubt that scientific
theorising is the most appropriate model of a central cognitive process.
Inferential comprehension, which we are claiming is also a central process,
differs from scientific theorising in a number of relevant respects. First,
although both processes are global in Fodor's sense, they operate on very
different time-scales. Because the construction and evaluation of a
scientific theory may take all the time in the world, the range of
hypotheses that can be considered, and the range of evidence that can be
taken into account, can be enormous, not just in theory but in practice. By
contrast, ordinary utterance comprehension is almost instantaneous, and
however much evidence might have been taken into account, however
many hypotheses might have been considered, in practice the only
Non-demonstrative inference
67
evidence and hypotheses considered are those that are immediately

accessible.
In the second place, the data for scientific theorising come from nature,
which is not actively involved in helping humans build correct scientific
theories. By contrast, the data for the comprehension process come from a
helpful source. People would not communicate ostensively if they did not
want their communicative intentions to be recognised, and they devise
their stimuli accordingly. Moreover, while it is quite conceivable that it is
beyond the power of humans to construct a fully adequate scientific
theory, successful inferential comprehension is demonstrably wirhin the
grasp of the normal intellect. Verbal comprehension in particular, with its
well-described linguistic stimuli and relatively clear criteria of success, is
much more amenable to investigation than scientific theorising. Precisely
because it is a central process, and is not a separate, purpose-built ability
like the visual or grammatical abilities, an adequate account of inferential
comprehension should shed light on other central processes, about which,
as Fodor rightly emphasises, so little is so far known.
As we will show in chapter 3, the fact that verbal comprehension is
almost instantaneous, and is achieved with the active help of the
information source, i.e. the speaker, makes the hearer's choice of a context
from the whole of conceptual memory more amenable to study.
However, the richness of the accessible information is only one of the two
main obstacles to investigation of the central cognitive processes. The
other has to do with the nature of the inference processes that this
information undergoes. Although logic provides us with several models of
demonstrative inference, it is agreed that the inference processes involved
in comprehension are non-demonstrative. While it is generally assumed
that non-demonstrative inference must be based on inductive rules of
some kind, there is no well-developed system of inductive logic that
would provide us with a plausible model of the central cognitive
processes.
Moreover, humans may be capable of controlling more than one
technique for performing non-demonstrative inference. A scientist selfconsciously applying explicit standards of confirmation to each piece of
available evidence may well be using a quite different system from those
we all use - scientists included - in making spontaneous, instantaneous
and unconscious inferences about the movements of other vehicles while
driving a car, about what some appetising food might taste like, or about a
speaker's communicative intention. Here, we are only concerned with
spontaneous non-demonstrative inference, which we take to be of more
general psychological significance than the painstakingly acquired inferential techniques of the scientist.
Even the claim that a proper model of spontaneous non-demonstrative
Inference
68
inference is to be found in a system of inductive logic is open to question.
Inference is the process by which an assumption is accepted as true Or
probably true on the strength of the truth or probable truth of other
assumptions. It is thus a form of fixation of belief. There are other forms:
perception, for instance, is a process by which an assumption is accepted
as true or probably true on the strength of a non-conceptual cognitive
experience. Demonstrative inference, the only form of inference that is
well understood, consists in the application of deductive rules to an initial
set of premises. There is thus a temptation to think of non-demonstrative
inference as the application of non-deductive infere11fe rules. However,
this temptation is based on analogy rather than argument. In fact, there is
reason to doubt that spontaneous non-demonstrative inference, as
performed by humans, involves the use of non-deductive inference rules.
Deductive inference rules generate all the interesting conclusions
logically implied by a set of premises. 1 It is generally recognised that
non-demonstrative inference rules cannot be expected to generate all the
interesting conclusions non-demonstratively supported by a set of
premises. For instance, the theory of relativity could not have been
generated by applying inference rules to the results of Eddington's
experiment. Instead, the process of reaching valid non-demonstrative
conclusions is generally broken down into two distinct stages: hypothesis
formation and hypothesis confirmation. Eddington's experiment provided the first empirical confirmation of Einstein's theory, but did not in
any sense imply it. Hypothesis formation, it is argued, is a matter of
creative imagination; hypothesis confirmation, on the other hand, can be
seen as a purely logical process governed by inference rules.
The function of inference rules is to guarantee the logical validity of the
inferences they govern. In a valid demonstrative inference, the application
of deductive rules to tr,ue premises guarantees the truth of the conclusions.
Similarly, iljl a valid non-demonstrative inference, hypothesis confirmation could be seen as governed by logical rules. These confirmation rules
might apply jointly to the premises, or 'evidence', and the tentative
conclusions, or 'assumptions', and assign a degree of confirmation to the
assumptions on the basis of the evidence. It is tempting to move from these
logical considerations to psychological speculation.
Humans are rather good at non-demonstrative reasoning; otherwise the
species would be extinct. This might be because they have logical rules
which constrain the confirmation of assumptions in the way just
described. However, this is not much of an explanation, since we have no
clear idea of what these rules might be. Also, at this level of vagueness,
other explanations are possible. For all we know, human inferential
successes might be attributable not so much to logical constraints on
confirmation as to cognitive constraints on hypothesis formation.
Non-demonstrative inference
69
The constraints on human conceptual systems might be such that the

only spontaneously ?evisable hypotheses are those ~hich, if false, are very
likely to be contradIcted by perceptually fixed belIefs. In fact, there are
independent reasons for assuming that human conceptual systems are so
constrained. For instance, no human language contains a word such as
'grue', the troublesome term invented by Nelson Goodman (1955) which
applies to anything which is either green and examined before some time t,
or blue and examined after t. If t is the year 3000 and all the emeralds you
have ever seen have been green, then you have evidence that all emeralds
are grue, and therefore that all emeralds examined after year 3000 will be
blue. Moreover, the hypothesis that all emeralds are grue rather than green
could not be falsified by any empirical evidence available before the year
3000. It is quite remarkable that natural language predicates prevent such
paradoxes from arising.
We are suggesting, then, that non-demonstrative inference, as spontaneously performed by humans, might be less a logical process than a
form of suitably constrained guesswork. If so, it should be seen as
successful or unsuccessful, efficient or inefficient, rather than as logically
valid or invalid. We would like to pursue this idea and make an even
stronger claim: that the only logical rules spontaneously accessible to the
human mind are deductive rules. Deductive rules, we will argue, play a
crucial role in non-demonstrative inference. Of course, the validity of a
deductive inference does not guarantee the validity of the overall
non-demonstrative inference of which it forms a part. Human spontaneous non-demonstrative inference is not, overall, a logical process.
Hypothesis formation involves the use of deductive rules, but is not
totally governed by them; hypothesis confirmation is a non-logical
cognitive phenomenon: it is a by-product of the way assumptions are
processed, deductively or otherwise.
By its very definition, a non-demonstrative inference cannot consist in a
deduction. Many authors seem to make the much stronger and unwarranted hypothesis that a non-demonstrative inference cannot contain a
deduction as one of its sub-parts. The recovery of implicatures, for
example, isa paradigm case of non-demonstrative inference, and it is
becoming a commonplace of the pragmatic literature that deduction plays
little if any role in this process. Leech (1983: 30-1) claims that the process
by which implicatures are recovered 'is not a formalised deductive logic,
but an informal rational problem-solving strategy', and that 'all implicatures are probabilistic'. Levinson (1983: 115-16) suggests that in certain
respects implicatures 'appear to be quite unlike logical inferences, and
cannot be directly modelled in terms of some semantic relation like
entailment'. Bach and Harnish (1979: 92-3) argue that the form of
inference by which implicatures are recovered 'is not deductive but what
70
Inference
might be called an inference to a plausible explanation'. Brown and Yule

(1983: 33) say, more generally,
It may be the case that we are capable of drawing a specific
conclusion ... from specific premises ... via deductive inference, but
we are rarely asked to do so in the everyday discourse we encounter.
. . . We are more likely to operate with a rather loose form of
inferencing.
Similar views are expressed by de Beaugrande and DressIer (1981: 93-4):

Humans are evidently capable of intricate reasoning processes that
traditional logics cannot explain: jumping to conclusions, pursuing
subjective analogies, and even reasoning in the absence of knowledge .... The important standard here is not that such a procedure is
logically unsound, but rather that the procedures work well enough
in everyday affairs.
However, those pragmatists who express scepticism about the role of
deductive reasoning in comprehension generally have little positive to say
about the nature of the inference processes involved. Bach and Harnish
comment, (1979: 93),
Our empirical thinking in general is rife with generalizations and
inference principles that we are not conscious of when we use them,
if we are conscious of them at all. It would take us well beyond
present-day cognitive psychology to speculate on the details of any
of this. Whatever these processes are, whatever activates them,
whatever principles or strategies are involved, they work, and work
well.
But the fa'Ct that these procedures work well enough in everyday utterance
comprehension does not absolve us from saying what they are. If
anything, the lack.of any existing framework for describing them should
make us more, not less interested in their nature. 2
Pragmatic theory in general is condemned to vagueness if it says
nothing more about the inference processes involved in comprehension
than that they are non-demonstrative, a purely negative characterisation.
Moreover pragmatic theories in which some notion of relevance plays a
role (i.e. most pragmatic theories) need some account of nondemonstrative inference for a second reason: one common way of
achieving relevance consists in providing the addressee with evidence
Forms, attitudes, assumptions
71
which bears on some assumption of his. Consider the following dialogue,

for instance:
(1) Peter: According to the weather forecast, it's going to rain.

Mary (standing at the window): It certainly looks like it.
Mary's remark does not prove that it will rain, but it confirms Peter's
belief and achieves relevance thereby. Relevant information, we have
suggested, is information that modifies and improves an overall representation of the world. Confirmation of an assumption by a nondemonstrative inference process is a case in point.
It seems, then, that the comprehension process may involve confirmation of assumptions in two quite different ways, or on two quite different
levels. On the one hand, as we saw in chapter 1, understanding a piece of
ostensive behaviour involves constructing and confirming a hypothesis
about the communicator's informative intention. On the other, as we
have just suggested, the most relevant effect of ostension may be to
confirm some previous assumption of the audience. A clearer account of
non-demonstrative inference processes should thus shed light on the role
of relevance in both communication and cognition: on what is involved in
constructing and confirming a hypothesis about the communicator's
intentions; on what it is for a representation of the world to be modified
and improved; and on the relation between the two.
We will suggest an approach to non-demonstrative inference which we
hope will go some way towards resolving these issues. But first we must
specify the domain of spontaneous non-demonstrative inference and say a
little more about what it is to have an overall representation of the world.
2 Logical forms, propositional attitudes and factual assumptions
Following Fodor (1983), we see the mind as a variety of specialised

systems, each with its own method of representation and computation.
These systems are of two broad types. On the one hand there are the input
systems, which process visual, auditory, linguistic and other perceptual
information. On the other hand there are the central systems, which
integrate information derived from the various input systems and from
memory, and perform inferential tasks. b
We assume that each input system has its own method of representation
and computation, and can process only information in the appropriate
representational format. Auditory perception can process only acoustic
information, and the processes involved in auditory perception differ
from those involved in olfactory perception, etc. One of the functions of
72
Inference
the input systems is to transform 'lower level' sensory representations into

'higher level' conceptual representations, which are all in the same format
regardless of the sensory modality from which they derive. It is because
they operate over such modality-neutral conceptual representations that
the central processes can integrate and compare information derived from
the various input systems and from memory.
The fact that many central processes are inferential puts an important
constraint on the conceptual representation system. Conceptual representations must have logical properties: they must be capable of implying
or contradicting one another, and of undergoing deductive rules. However, not all properties of a conceptual representation are logical properties.
A conceptual representation is both a mental state and a brain state. 3 As a
mental state, it can have such non-logical properties as being happy or sad.
As a brain state, it can have such non-logical properties as being located in
a certain brain at a certain time for a certain duration. Let us abstract away
from all these non-logical properties, and call the remaining logical
properties of a conceptual representation its logical form. It is in virtue of
its logical form that a conceptual representation is involved in logical
processes and enters into relations such as contradiction or implication
with other conceptual representations.
A logical form is a well-formed formula, a structured set of constituents,
which undergoes formal logical operations determined by its structure. ~
we have already said, what distinguishes logical operations from other
formal operations is that they are truth-preserving: a deduction from a
true representation P yields a true representation Q. By contrast, deleting
the first constituent of a representation, say, is a formal operation but not a
logical one. Given this relationship between truth and logic, it might seem
that only a conceptual representation which is capable of being true or
false can have a logical form. We see things rather differently. In essence,
we will argue that for a representation to be amenable to logical
processing, all that is necessary is for it to be well formed, whereas to be
capable of being true or false, it must also be semantically complete: that
is, it must represent a state of affairs, in a possible or actual world, whose
existence would make it true. We take it that an incomplete conceptual
structu~e can nevertheless be well formed, and can undergo logical
processmg.
Let us say that a logical form is propositional if it is semantically
complete and therefore capable of being true or false, and nonpropositionalotherwise. A formal example of a non-propositionallogical
form is a predicate calculus formula containing a free variable: this may be
syntactically well formed without being fully propositional. A psychological example of a non-propositional logical form is the sense of a
sentence. Given that 'she' and 'it' in (2) below do not correspond to
Forms, attitudes, assumptions
73
definite concepts, but merely mark an unoccupied place where a concept

might go, sentence (2) is neither true or false:
(2) She carried it in her hand.

In spite of its non-propositionality, (2) has obvious logical properties. For
instance, it implies (3), which is equally non-propositional, and it
contradicts (4), which is, or can be understood as, propositional:
(3) She held something in her hand.

(4) No one ever carried anything.
Incomplete logical forms play an important role in cognition. In the
first place, we will argue that they may be stored in conceptual memory as
assumption schemas, which can be completed into full-fledged assumptions on the basis of contextual information. In the second place, as we
have just seen, the sense of a sentence is often an incomplete logical form.
We will show in chapter 4 that when a natural-language sentence is
uttered, the linguistic input system automatically decodes it into its logical
form (or in the case of an ambiguous sentence a set of logical forms), which
the hearer is normally expected to complete into the fully propositional
form that the speaker intended to convey. 4
However, while non-propositionallogical forms may play an important role in intermediate stages of information processing, only fully
propositional forms represent definite states of affairs. It is these that
constitute the individual's encyclopaedic knowledge, his overall representation of the world.
The mind does not just construct and store logical forms: it entertains
them in different ways. A philosopher would say that the mind is capable
of different propositional attitudes; a cognitive psychologist might say
that different representations are processed and stored in different ways.
For instance, a propositional form can be entertained as a description of an
actual state of affairs, as a description of a desirable state of affairs, or as a
good rendering (e.g. a summary) of some other representation. An
individual encyclopaedic memory consists not merely of a stock of
conceptual representations but of a stock of representations with propositional or non-propositionallogical forms, entertained in different ways, as
objects of different attitudes such as belief and desire.
Utterances, too, convey different attitudes to the representations they
express. Some basic attitudes are conveyed by syntactic means, in
particular by the mood of the verb. In English, for instance, there is a
correspondence between indicative mood and the attitude of belief, and
between imperative mood and the attitude of desire. Other attitudes are
expressed by lexical means: that is, the speaker's attitude is made explicit
in a main clause, as in (5), or a parenthetical clause, as in (6):
-74
Inference
(5) I wish that P.

(6) P, I suppose.
The attitudes which can be expressed lexically are much more numerous
and varied than those which can be expressed syntactically. Arguably,
similar alternatives exist in the conceptual representation system, the
language of thought. That is, there may be one or two basic formats for
distinguishing, say, the attitudes of belief and desire, together with a range
of conceptual resources for expressing or recording a wide range of further
attitudes-. We would like to pursue this possibility.
Let us assume that there is a basic memory ltore 5 with the following
property: any representation stored in it is treated by the mind as a true
description of the actual world, a fact. What this means is that a
fundamental propositional attitude of belief or assumption is pre-wired
into the very architecture of the mind. As a r.esult, a representation can be
entertained as an assumption without the fact that it is an assumption
being explicitly expressed. Such basic assumptions, entertained as true
descriptions of the world, but not explicitly represented as such, we will
call factual assumptions.
The human internal representation system is clearly rich enough to
allow for second-order representations of representations. In other
words, the language of thought acts as its own meta-Ianguage: we are
capable not only of entertaining assumptions but also of thinking about
them and about other representations. It is thus possible not merely to
entertain the belief that P, but to represent to oneself the fact that one
believes that P, or that someone else believes that P, or that one believes
that someone else believes that P, and so on. The belief or assumption
that P can thus be held in two different ways: either as the basic factual
assumption that P, or as the factual assumption I believe that P.6
Conceivably, the attitude of desire might parallel the attitude of belief in
having its own basic memory store or storage format. This would mean
that desire, like belief, was pre-wired into the architecture of the human
cognitive system. It would follow that the desire that P could be held in
two different ways: either as the basic desire that P, or as the factual
assumption I desire that P. Alternatively, there might be only a single
basic memory store, the one used for factual assumptions. In that case
desires could play a cognitive role only by being represented in factual
assumptions of the form I desire that P.
As far as we can see, factual assumption is the only obvious case of an
attitude marked off by a special form of storage, and desire is the only
other plausible case. It seems unlikely that other propositional attitudes e.g. doubting that P, regretting that P, fearing that P, pretending that Phave their own basic memory stores. If we are right, these attitudes can
75
play a cognitive role only via factual assumptions of the form It is dubious
that P, I regret that P, I am afraid that P, and so on.
A representation of the world, then, may be regarded without too much
oversimplification as a stock of factual assumptions, some basic, others
expressing attitudes to embedded propositional or non-propositional
representations. Factual assumptions are the domain par excellence of
spontaneous non-demonstrative inference processes. Each newly
acquired factual assumption is combined with a stock of existing
assumptions to undergo inference processes whose aim, we have suggested, is to modify and improve the individual's overall representation of
the world. 7
When a representation is stored not as a basic factual assumption but by
being embedded under an expression of attitude, it is often processed in a
self-conscious, non-spontaneous way. This is true of representations used
in problem-solving tasks of the kind familiar from experimental psychology. It is true of speculatively held opinions, religious beliefs, or scientific
hypotheses. The largely conscious reasoning processes which these
indirectly held representations undergo are of great intrinsic interest, but
we see it as a mistake to extrapolate from them to the spontaneous and
essentially unconscious inference processes used in most ordinary thinking, and in particular in ordinary verbal comprehension.
The model of inferential communication and the notion of relevance we
are developing are not tied to any particular form of inference. We assume,
for instance, that the lengthy and highly self-conscious processes of
textual interpretation that religious or literary scholars engage in are
governed just as much by considerations of relevance as is spontaneous
utterance comprehension. However, in this book we want to focus on the
latter. Spontaneous inference plays a role even in scholarly interpretation,
whereas scholarly thinking is a rather exceptional human endeavour, even
for scholars. The study 'of spontaneous inference is thus a necessary
prerequisite to a proper investigation of all forms of human inference,
including inferential communication.
3 Strength of assumptions
Factual assumptions are entertained with greater or lesser confidence; we
think of them as more or less likely to be true. We do so consciously in two
main types of situation. First, we may have to choose between contradictory assumptions, as when I assumed that Bob would be out of town, and
now I assume that I am seeing him walking down the street: of the two
assumptions, the one I regard as more likely to be true will displace the
other. Second, we may have to choose between different courses of action,
76
Inference
as when I want to buy some petrol, and I assume that both the petrol
station up the street and the one down the street are open: if I regard the
assumption that the station down the stt:eet is open as more likely to be
true, that is where I will go.
When our more confident assumptions are those which are in fact more
likely to be true, we tend to make the right choices of assumptions and
courses of action. In other words, the adequacy of our representation of
the world depends not only on which assumptions we hold, but also on
our degree of confidence in them: an adequate representation is one in
which there is a good match between the assumRtions we regard as well
confirmed and those that actually are well confirmed. Improvements in
our representation of the world can be achieved not only by adding
justified new assumptions to it, but also by appropriately raising or
lowering our degree of confidence in them, the degree to which we take
them to be confirmed.
'Confirmation' is a term taken from a relatively undeveloped branch of
logic (which itself took it from commonsense psychology). How should it
be adapted for use in cognitive psychology? Two very different answers
can be given. On one approach, the logical concept of confirmation makes
a good psychological concept as it stands; the logician's system is, on the
whole, an -adequate psychological model. All that is needed is to replace
the objective notion of confirmation with some subjective analogue: for
example, a system assigning subjective probability values to representations. Call this the logical view.
On another approach, the logical concept of confirmation should be not
adapted but rejected. The ability to judge an assumption as more or less
likely to be true is to be explained not in terms of a system which assigns
subjective probability values to assumptions, but in terms of a non-logical
property of assumptions: what, metaphorically, we will call their strength.
Call this the non-logical, or functional, view. This is the approach that we
would like to pursue.
According to the logical view, every factual assumption consists of two
representations. The first is a representation of a state of affairs - for
instance (7a); the second is a representation of the confirmation value of
the first representation - for instance (7b):
(7) (a) Jane likes caviar.
(b) The confirmation value of (a) is 0.95.
How are these two representations arrived at? The first, so the story goes,
is the output of a non-logical cognitive process of assumption formation.
The second is the output of a process of logical computation which takes
as input the assumption to be confirmed, on the one hand, and the
available evidence, on the other. When new evidence becomes available, a
77
new logical computation may take place, as a result of which the

confirmation value of an assumption may be raised or lowered. On this
view, the notion of a confirmation value is basic, and the strength of an
assumption - if this is worth talking about at all - is determined by its
confirmation value.
According to the functional view, a factual assumption consists of a
single representation, such as (7a). The strength of this assumption is a
result of its processing history, and cannot be accounted for in terms of the
logical concept of confirmation. Understood in this way, the strength of
an assumption is a property comparable to its accessibility. A more
accessible assumption is one that is easier to recall. For instance, for most
of our readers who assume (8) and (9), (8) is more accessible than (9):
(8) Cairo is the present capital of Egypt.
(9) Thebes was the capital of Egypt under the 20th dynasty.
Clearly, this difference is not to be explained by an appeal to two
second-order representations assigning different degrees of accessibility
to (8) and (9). A more plausible claim is that, as a result of some kind of
habituation, the more a representation is processed, the more accessible it
becomes. Hence, the greater the amount of processing involved in the
formation of an assumption, and the more often it is accessed thereafter,
the greater its accessibility.
Similarly, the initial strength of an assumption may depend on the way
it is acquired. For instance, assumptions based on a clear perceptual
experience tend to be very strong; assumptions based on the acceptance of
somebody's word have a strength commensurate with one's confidence in
the speaker; the strength of assumptions arrived at by deduction depends
on the strength of the premises from which they were derived. Thereafter,
it could be that the strength of an assumption is increased every time that
assumption helps in processing some new information, and is reduced
every time it makes the processing of some new information more
difficult; On this account, these variations in strength are neither the
object nor the output of a special logical computation. They arise, rather,
as by-products of various cognitive processes, deductive and nondeductive.
Here is an informal illustration. J ane told me herself that she likes caviar,
and I have no reason to doubt her word; so I hold assumption (7a) quite
strongly. On one occasion, I observe Jane eating caviar with a big smile;
my assumption helps me understand this fact and becomes even stronger
as a result. On another occasion, though, I seeJane turning down an offer
of caviar; accessing my assumption that she likes caviar not only does not
help, but makes this fact more difficult to understand; as a result, the
assumption becomes weaker. At no point do I process a representation of
78
Inference
the confirmation value of (7a). The strength of (7a) is established and
varies as a by-product of other processes, and need not be represented at
all in order to exist and vary.
However, this is not the whole story. Functional properties of
representations, such as accessibility or strength, need not be represented
in the mind in order to exist, vary and affect cognitive processes, but they
can be represented. People have intuitions about the accessibility of
different assumptions; we appealed to such intuitions about example
(8)-(9) above. Similarly, people have intuitions about the strength of their
assumptions. They may express these intuition~ in different ways. They
may say such things as
(10) I am quite certain that Jane likes caviar.
(11) I firmly believe that J ane likes caviar.
(12) What I have seen confirms my assumption thatJane likes caviar.
(13) It seems more probable to me that Jane likes caviar than that she likes
oysters.
Behind these forms of expression lies a tacit hypothesis. We take for
granted that there is a good match between the strength of our
assumptions and the likelihood that they are true. That is, we trust our
cognitive mechanisms to strengthen or weaken our assumptions in a way
that is epistemologically sound: we trust our representation of the world
to be adequate in this respect, as in others. As a result, intuitions about the
strength of our assumptions are expressed as intuitions about their degree
of confirmation. Such intuitions are assumptions about assumptions, and
can be processed as such. More often than not, they are assumptions about
changes in the strength of a single assumption, as in (12), or about the
relative strength of pairs of assumptions, as in (13). In some cases, it seems
plausible that such assumptions play a genuine cognitive role: for instance,
in conscious attempts to resolve a contradiction by working out which of
the contradictory assumptions is the more likely to be true.
According to the logical view, the soundness of our assumptions
depends on our ability to carry out a computational check on the
confirmation value of each assumption. According to the functional view,
the soundness of our assumptions - to the extent that they are sound depends on our having cognitive mechanisms so attuned to the world we
live in that the strength of our assumptions tends to match the likelihood
that they are true. On the logical view, a representation of the confirmation value of an assumption is an aspect of that assumption; it is the
outcome of a logical process which every assumption undergoes. On the
functional view, a representation of the degree of confirmation of an
assumption is another assumption; it is generally the product of an
intuition about one of the effects of the processing history of that
79
assumption. Assumptions can exist without their degree of confirmation

ever being represented.
Neither the logical nor the functional view has been developed to a
point where they could be effectively tested. Nonetheless, they are
different enough for some empirical comparisons to be made. We will try
to show that these comparisons fayour the functional view.
Our conscious intuitions about the strength of assumptions yield only
the grossest kinds of absolute judgement. We may think of a given
assumption as certain (true), very strong (very well confirmed), strong
(well confirmed), weak (poorly confirmed), but the boundaries between
these categories are fuzzy. Unless we have been trained in inductive logic,
we have no sub-categories to express more precise absolute judgements.
On the other hand, we can often make much finer comparative
judgements. For instance, we may be aware that some new piece of
evidence has strengthened an assumption even though it remains in the
same absolute category: say it was strong before and is still strong, but we
think of it as stronger than before. However, not all comparative
judgements are equally easy. If two assumptions are utterly unrelated, and
if they belong to the same gross category, say strong, they become well
nigh impossible to compare. Consider (14) and (15):
(14) Jane likes caviar.
(15) There are more Indian restaurants than Chinese restaurants
Chelsea.
In
Someone who quite strongly believes (14) and (15) might find it difficult
or impossible to answer the question, 'Do you think it is more likely that
Jane likes caviar than that there are more Indian restaurants than Chinese
restaurants in Chelsea?'
It seems to us implausible that humans might have a system for
computing and representing the strength of assumptions which is both
wholly unconscious and radically more sophisticated than anything that is
reflected in their conscious intuitions. We therefore reject the possibility
that an individual might unconsciously assess cpnfirmation values
through the kind of numerical computations suggested by logicians, when
he is incapable of doing so consciously. 8 We conclude, more generally,
that the strength of an assumption cannot be quantitatively assessed: in the
terms of Carnap (1950), it is a comparative rather than a quantitative value.
In his 1950 treatise on subjective probability (another term for
estimated degree of confirmation), Carnap contrasts classificatory, comparative and quantitative concepts along the following lines.
A classificatory concept sets up a necessary and sufficient condition for
class membership. For example, an integer either is prime (if it can be
evenly divided by no other whole number than itself and 1), or it is not.
80
Inference
A comparative concept is one that figures in comparative judgements.
For example, some things feel warmer than others, some sounds seem
louder than others, some foods taste nicer than others, and so on. Some
classificatory concepts have comparative counterparts, but not all of them
do. Caffeinated is both a classificatory concept (a substance either does or
does not contain caffeine) and a comparative one (some substances contain
more caffeine than others). On the other hand, prime does not have a
comparative interpretation: numbers cannot be more or less prime.
Carnap's third type of concept, the quantitative concept, is one that
figures in numerical comparisons. For example: distance is a quantitative
concept because we can not only say that London is further from
Edinburgh than it is from Oxford, but also measure the distance in miles
or kilometres, and hence say how much further London is from
Edinburgh than it is from Oxford. However, as Carnap points out, not
every comparative concept has a quantitative counterpart. Though we
might know that one food tastes better than another, there is no obvious
way of measuring how good a food tastes, and therefore of measuring how
much better one food tastes than another.
The existence of an objective numerical scale makes it easy to formulate
precise absolute judgements and to compare unlike objects: say the ages of
a child and a car, or the distance from Trafalgar Square to Buckingham
Palace and from the foot of Mount Everest to the top. Where no numerical
scale exists, absolute judgements become gross, and comparison of unlike
objects becomes much harder. For example, someone might be able to say
that one champagne tastes better than another, or that one caviar tastes
better than another, but be quite unable to say whether or not a certain
champagne tastes better than a certain caviar.
Even where an objective numerical scale exists, it does not follow that
some internal analogue of it is used in mental comparison. For example,
when a suitcase feels heavier the longer it is carried, this feeling is
presumably not based on any assumption that the suitcase is actually
gaining ounces or pounds as the journey proceeds. Similarly, comparisons
of the warmth of unlike objects, such as a liquid with a solid, or a solid
with a gas, are much harder than comparisons of the warmth of like
objects. This strongly suggests that ordinary judgements of warmth are
not based on an internal analogue of a temperature scale.
Difficulties in comparing unlike objects are hard to explain on the
assumption that a numerical scale is being used. Where such difficulties
arise, it seems more reasonable to assume that what is being used is not a
numerical scale but rather a heuristic (based for instance on matching
procedures), which applies only to like objects: comparing, for example,
the taste of several samples of caviar, the warmth of various liquids, the
strength of several assump_tions of related content.
81
The performance of humans in comparing the strength of unlike

assumptions is thus a powerful indication that strength, as a basic
psychological concept applied to assumptions, is comparative rather than ..
quantitative. This fits much better with the functional view than with the
logical view. On the functional view, judgements and comparisons of
strength are introspective intuitions, just like judgements and comparisons of taste, pain or accessibility, all of which are clearly comparative. Of
course, this does not explain how these judgements and comparisons are
made; but it makes the problem of degrees of confirmation one among
many: no new problem is created. The logical view, on the other hand,
implies that each assumption is assigned an absolute quantitative confirmation value. This raises two new problems: since fine comparisons are
possible in Some cases, why aren't they always possible? And why are our
conscious absolute judgements of strength so gross?
Other arguments for the functional view could be derived from a study
of the logical mistakes that subjects make in evaluating degrees of
confirmation. 9 We are aware that neither these further arguments nor
those we have put forward are conclusive. What we hope to have shown so
far is that the functional view has some initial plausibility and may be
worth pursuing.
On the functional approach, the successes of human non-demonstrative
inference must be explained by appealing not to logical processes of
assumption confirmation, but to constraints on the formation and
exploitation of assumptions. Factual assumptio~s are acquired from four
sources: perception, linguistic decoding, assumptions and assumption
schemas stored in memory, and deduction. In the rest of this section we
will look briefly at how assumptions from these four sources are formed,
and at how they acquire an initial degree of strength.
Perceptual mechanisms assign to a sensory stimulus a conceptual
identification of that stimulus, e.g.
(16) This is an orchid.
(17) The doorbell is ringing.
(18) The pavement is wet.
Under normal conditions of perception, these elementary descriptions of

stimuli become strong assumptions. That these assumptions are generally
correct is due to the fact that human perceptual mechanisms are the
outcome of a long b!910gical evolutionary process, and are well adapted to
the task.
The linguistic input mechanisms assign to a particular type of sensory
stimulus a logical form. We have seen that the logical forms recovered by
decoding fall short of being fully propositional, and that even when
completed into propositional forms they fall short of being factual
82
Inference
assumptions. However, the propositional form arrived at by completing

the logical form of an uttered sentence can be integrated by a standard
procedure into an assumption about what the speakersaid. For instance, if
Peter is heard uttering (19) at time t, his utterance will be decoded as
having the logical form of sentence (20), which can be completed to yield
the propositional form (21), which can in turn be integrated into the
assumption schema in (22) to yield assumption (23):
(19)
(20)
(21)
(22)
(23)
[alha:vJhedelk]
I have a headache.
Peter has a headache at time t.
Peter says that - - .
Peter says that Peter has a headache at time t.
The processes by which assumptions such as (23) are formed and

exploited, and the empirical adequacy of these assumptions, will be
discussed extensively in chapter 4.
Conceptual memory is a huge repertory of assumptions such as
(24)-(28):
The car is in the garage.
Larry is a philosopher.
Orchids are rare flowers.
When the outside temperature is below five degrees centigrade, the
pond is frozen.
(28) Philosophers are entertaining.
(24)
(25)
(26)
(27)
We assume that memory also contains assumption schemas, i.e. logical

forms which can be completed to yield propositional forms in the format
appropriate for factual assumptions. lo Thus, assumption schema (29)
might be completed to yield assumption (30), and assumption schema (31)
might be completed to yield assumption (32):
(29)
(30)
(31)
(32)
The outside temperature is - - degrees centigrade.

The outside temperature is minus six degrees centigrade.
- - is (a bachelor/married/divorced/a widower).
Larry is a bachelor.
It also appears that when available assumptions correspond to a certain

schema, related schemas are used to derive further assumptions. For
instance, when assumptions of the form (33) are made, it seems that
assumptions of the form (34) or (35) are standardly considered:
(33) If P then Q.
(34) If (not P) then (not Q).
(35) If Q, then (Q because P).
83
Thus, the formation of assumption (36) would standardly lead to a

consideration of assumptions (37) or (38):
(36) If Fido is pleased, then he wags his tail.
(37) If Fido is not pleased, then he does not wag his tail.
(38) If Fido wags his tail, then he does so because he is pleased.
Assumptions retrieved from memory come with a certain degree of
strength. Assumptions constructed by completing assumption schemas
come with an initial plausibility which may make them worth processing;
their subsequent strength depends on their subsequent processing history.
Given a set of assumptions as premises, further assumptions can be
derived as conclusions of a deductive process. For instance, from (16) and
(26), assumption (39) can be derived:
(16) This is an orchid.
(26) Orchids are rare flowers.
(39) This is a rare flower.
Similarly, (40) can be derived from assumptions (25) and (28):
(25) Larry is a philosopher.
(28) Philosophers are entertaining.
(40) Larry is entertaining.
Assumption (41) can be derived from (27) and (30):
(27) When the outside temperature is below zero centigrade, the pond is
frozen.
(30) The outside temperature is minus six degrees centigrade.
(41) The pond is frozen.
We will argue that the formation of assumptions by deduction is the key
process in non-demonstrative inference, and show how new assumptions
inherit their strength from the assumptions used in deriving them. But
first we must give some thought to the deductive process itself.
4 Deductive rules and concepts
The deductive processing of information has much of the automatic,

unconscious, reflex quality of linguistic decoding and other input
processes. What distinguishes the deductive system from the input
systems is that it applies to conceptual rather than to perceptual
representations: that is, to representations with a logical or propositional
form. What distinguishes it from other central systems is the type of
computation it performs.
84
Inference
Deductive arguments can be looked at from a syntactic (computational)

or a semantic point of view. Let us say that a semantic relation of
entailment holds between two assumptions P and Q if and only if every
conceivable state of affairs which would make P true would also make Q
true: that is, if and only if there is no conceivable state of affairs in which P
would be true and Q false. The relation is a semantic one in that it involves
reference to the states of affairs which particular assumptions represent:
that is, the states of affairs which constitute their semantic interpretations.
By this definition, (42) entails (43), since there is no conceivable state of
affairs in which (42) would be true and (43) false:,
(42) Apples grow in orchards and grapes grow in vineyards.
(43) Apples grow in orchards.
Let us say that a syntactic relation of logical implication holds between

two assumptions P and Q with respect to a certain deductive system if and
only if one is deducible from the other by the deductive rules of that
system. A deductive rule is a computation which applies to assumptions in
virtue of their logical form. Logical implication is a syntactic relation in
that it holds purely in virtue of the formal properties of assumptions, and
involves no reference to their semantic properties. For example, most
standard logics have a rule of and-elimination which applies to assumptions of the form (P and Q) to yield conclusions of the form P, or
conclusions of the form Q. In a system with such a rule, (42) logically
implies (43).
There is a necessary connection between logical implication and
entailment in at least the following sense: the notion of a deductive rule
itself cannot be properly explicated without appeal to the semantic notion
of entailment. Formally, a deductive rule is a computation like any other.
What distinguishes it from other computations is the fact that it is a
truth-preserving operation: that is, when it applies to an assumption, the
conclusion it yields stands in a semantic entailment relation to the premise.
Hence, all logical implications are also entailments. However, the reverse
relation does not necessarily hold.
Although most standard logics in principle aim at completeness, which
means that they aim at a system in which every assumed entailment is also
a logical implication, in practice they ignore entailments which do not
hinge on the meanings of a small class of 'logical particles' such as and, or,
not, some and all. For example, there is no deductive rule in standard logics
which would derive (45) from (44), even though (44) entails (45):
,
(44) All bachelors are happy.

(45) All unmarried men are happy.
Nor is there any reason why there should be from the purely logical point
85
of view. There is no contradiction in the idea of an incomplete logic, as

there is in the idea of a logic whose rules are not at least partial
reconstructions of assumed entailment relationships.
The question of whether humans have deductive rules as part of their
basic mental equipment, and if so, which rules they have, is of no
particular interest to pure logicians. Logicians are concerned with the
nature of conceivable deductive systems, whether psychologically realised
or not. However, it is a question of considerable interest for cognitive
psychology in general and pragmatic theory in particular. We assume, as
do most others working in the area, that there is a set of deductive rules
which are spontaneously brought to bear in the deductive processing of
information. This is an empirical assumption, which might be justifiable
along the following lines.
First, for any organism which represents the world in conceptual terms,
that is, in terms of a set of assumptions, a deductive system would effect an
important economy of storage. Given a set of deductive rules, the logical
implications of any set of assumptions would be recoverable from it by
means of the deductive rules, and would not need to be separately stored.
Second, for any organism interested in improving its conceptual representation of the world, it would provide a tool not only for working out
the consequences of adding a new assumption to an existing representation of the world, but for guaranteeing the accuracy of any conclusions
deduced from initially accurate premises. Third, for any organism
interested in the accuracy of its conceptual representation of the world, it
would provide a tool for exposing inconsistencies, and hence inaccuracies,
in any existing representation. No other system of inference with similar
powers has ever been developed with a degree of explicitness comparable
to that of a deductive systemY C
We assume, then, that at the heart of the human ability to perform
spontaneous demonstrative inference is a set of deductive rules: a set of
computations which take account of the semantic properties of assumptions only insofar as these are reflected in their form. 12 We have so far said
little about the form of assumptions, in virtue of which they undergo
deductive rules. In what follows, we will sketch an account which, though
speculative, is as far as we know compatible with the available empirical
evidence. Some of these speculations have the sole function of showing
that our general claims about verbal comprehension could in principle be
psychologically realised; others are more substantive.
It seems reasonable to regard logical forms, and in particular the
propositional forms of assumptions, as composed of smaller constituents
to whose presence and structural arrangements the deductive rules are
sensitive. These constituents we will call concepts. An assumption, then, is
a structured set of concepts.
86
Inference
Concepts, like the logical forms that contain them, are psychological
objects considered at a fairly abstract level. Formally, we assume that each
concept consists of a label, or address, which performs two different and
complementary functions. First, it appears as an address in memory, a
heading under which various types of information can be stored and
retrieved. Second, it may appear as a constituent of a logical form, to
whose presence the deductive rules may be sensitive. These functions are
complementary in the following sense: when the address of a certain
concept appears in a logical form being processed, access is given to the
various types of information stored in memory at that address.
The information that may be stored in memory- at a certain conceptual
address falls into three distinct types: logical, encyclopaedic and lexical.
The logical entry for a concept consists of a set of deductive rules which
apply to logical forms of which that concept is a constituent. The
encyclopaedic entry contains information about the extension and/or
denotation of the concept: that is, about the objects, events and/or
properties which instantiate it. The lexical entry contains information
about the natural-language counterpart of the concept: the word or phrase
of natural language which expresses it. On this approach, a conceptual
address is thus a point of access to the logical, encyclopaedic and linguistic
information which may be needed in the processing of logical forms
containing that address. We consider each type of entry in turn.
A logical entry consists of a set of deductive rules, each formally
describing a set of input and output assumptions: that is, a set of premises
and conclusions. Our first substantive claim is that the only deductive
rules which can appear in the logical entry of a given concept ate
elimination rules for that concept. That is, they apply only to sets of
premises in which there is a specified occurrence of that concept, and yield
only conclusions from which that occurrence has been removed.
Standard logics invariably contain such rules. For example, the standard
logical rule of and-elimination takes as input a single conjoined premise
and yields as output one of its constituent conjuncts:
(46) And-elimination
(a) Input: (P and Q)
Output: P
(b) Input: (P and Q)
Output: Q
That is, it applies only to premises containing a designated occurrence of
the concept and, and yields conclusions from which that occurrence has
been removed. The standard rule of modus ponendo ponens takes as input
a pair of premises, one a conditional and the other the antecedent of that
conditional, and yields as output the consequent of the conditional:
87
(47) Modus ponendo ponens

Input: (i) P
(ii) (If P then Q)
Output: Q
the concept if ... then, and yields conclusions from which that occurrence
has been removed. The standard rule of modus tollendo ponens takes as
input a pair of premises, one a disjunction and the other the negation of
one disjunct, and yields as output the other disjunct:
(48) Modus tollendo ponens

(a) Input: (i) (P or Q)
(ii) (not P)
Output: Q
(b) Input: (i) (P or Q)
(ii) (not Q)
Output: P
the concept or, and yields conclusions from which that occurrence has
been eliminated. We assume that some version of these rules is contained
in the logical entries for the concepts and, if . .. then and or, respectively.
Standard logics make a radical distinction between concepts such as
and, if. .. then, and or, which are regarded as proper logical concepts, and
concepts such as when, know, run, bachelor, which are considered
non-logical. Following another tradition, 13 we regard these other concepts
as also determining logical implications. Which concepts do or do not
have logical entries, which rules these entries contain, and which natural
classes concepts fall into from a cognitive point of view, are all matters
for empirical investigation. So far we have simply suggested, as an
empirical hypothesis, a general restriction on the form of logical entries.
These questions will be taken up in section 5.
The second type of entry attached to a concept, its encyclopaedic entry,
contains information about its extension and/or denotation: the objects,
events and/or properties which instantiate it. For example, the encyclopaedic entry for the concept Napoleon would contain a set of assumptions
about Napoleon, the encyclopaedic entry for the concept cat would
contain a set of assumptions about cats, and the encyclopaedic entry for
the concept argue would contain a set of assumptions about arguing.
Quite a lot of work has been done in the last ten or fifteen years on the
organisation of conceptual information in memory, and various models
have been proposed to describe what we are calling encyclopaedic entries.
These models are intended to answer questions about the structure of the
88
Inference
entries, the relations between the various types of assumption contained in
them, and the relations among the entries themselves. Many of the models
that have been proposed incorporate such notions as schema, frame,
prototype or script. 14
,
The idea behind these notions is that humans are disposed to develop
stereotypical assumptions and expectations about frequently encountered
objects and events. For example, I have an idea of a typical pet, which
includes dogs and cats but excludes elephants and spiders. It is thought
that these schematic assumptions and expectations are stored and accessed
as a unit or 'chunk', that they are highly accessi~le, and that they will be
used in default of any more specific information in processing utteran~es
about the associated objects or events. Thus, when I hear that my
.neighbour has bought a pet, I will assume that it is something like a dog or
a cat rather than an elephant or a spider, unless given specific information
to the contrary. We do not want to argue for or against any particular one
of these models. We share the basic hypothesis which is common to all of
them: in our terms, that encyclopaedic information contains not only
factual assumptions but also assumption schemas which an appropriate
context may convert into full-fledged assumptions.
Intuitively, there are clear-enough differences between encyclopaedic
and logical entries. Encyclopaedic entries typically vary across speakers
and times: We do not all have the same assumptions about Napoleon or
about cats. They are open-ended: new information is being added to them
all the time. There is no point at which an encyclopaedic entry can be said
to be complete, and no essential minimum without which one would not
say that its associated concept had been mastered at all. Logical entries, by
contrast, are small, finite and relatively constant across speakers and times.
There is a point at which the logical entry for a concept is complete, and
before which one would not say that the concept had been mastered at all.
Suppose, for example, that a child has not yet realised that X knows that P
implies P, and so uses know interchangeably with believe. We would say
that he had not yet mastered the concept. On the other hand, if he has
grasped this logical point but is unable to think of a single instance of
something he is prepared to call knowledge, we would regard this as a
failure of memory or experience (or a mark of philosophical potential)
rather than of understanding.
The distinction between logical and encyclopaedic entries corresponds
in many ways to the traditional distinction between analytic and synthetic
truths, which has been a notorious subject of dispute. However, our claim
is not so much that there is a fundamental difference between two types of
truth, two types of information content, as that information must be
representable in two different forms, and function in two different ways, if
89
successful communication is to take place. This claim seems to us

incontrovertible, at least within the current climate of research.
The whole framework of current cognitive psychology rests on a
distinction between representation and computation, of which our
distinction between encyclopaedic assumptions and deductive rules is a
special case. The information in encyclopaedic entries is representational:
it consists of a set of assumptions which may undergo deductive rules. The
information in logical entries, by contrast, is computational: it consists of
a set of deductive rules which apply to assumptions in which the
associated concept appears. It is not that the same item of information may
not be stored now in one form, now in another, or in both forms
simultaneously. For example, rules (46a-b) above (the rules of and-.
elimination) could be put into the propositional form in (49), and it is quite
possible that, on reflection, one might produce (49) as a description of
one's own computational practice:
(49) From a conjunction as premise it is valid to infer either conjunct as
conclusion.
The point is that representation and computation are two formally distinct
and complementary processes, neither of which can exist without the
other, and both of which are necessary for comprehension to take place. It
is this distinction that is reflected in our distinction between logical and
encyclopaedic entries.
It is also common practice in current cognitive psychology to distinguish between the content of a particular item of information (in this case an
assumption) and the context in which it is processed. Our suggestion is
that, broadly speaking, the content of an assumption is constrained by the
logical entries of the concepts it contains, while the context in which it is
processed is, at least in part, determined by their encyclopaedic entries.
Again, the point is not that the same piece of information could not
function, now as part of the content of an assumption, now as part of the
context in which it is processed; indeed there is reason to think that just
such overlaps do occur. The point is that in order to make sense of the
claim that assumptions are processed in a context - and in particular of our
claim that the relevance of an assumption is analysed in terms of the
modification that it brings to the context in which it is processed - it must
be possible in principle to distinguish between the content of an
assumption and its context. This distinction is reflected in our distinction
between logical and encyclopaedic entries.
The distinction between logical and encyclopaedic entries is thus quite
fundamental to our framework, and indeed to any framework that seems
90
Inference
to us remotely plausible. However, it is primarily a formal and functional
distinction, and does not necessarily imply that there are two fundamentally different kinds of truth.
The third type of entry for a concept, its lexical ,entry, contains
information about the natural-language lexical item used to express it. We
assume that this entry includes the sort of syntactic and phonological
information that would be contained in the lexical entry for that item in a
generative grammar: information about its syntactic category membership and co-occurrence possibilities, phonological structure, and so
(j
,
on.
The fact that concepts have both logical and lexical entries provides a
point of contact between input and central processes, between the
linguistic input system and the deductive rules of the central conceptual
system. Recovery of the content of an utterance involves the ability to
identify the individual words it contains, to recover the associated
concepts, and to apply the deductive rules attached to their logical entries.
We assume, then, that the 'meaning' of a word is provided by the
associated concept (or, in the case of an ambiguous word, concepts). This
allows us to maintain a somewhat ecumenical view of lexical semantics.
Most theories of lexical semantics assume that all words, with the possible
exception of proper names, have meanings of the same format. They then
differ as to what this universal format is. We recognise the possibility that
different words may have meanings of different formats.
A classical view is that the meaning of a word is provided by a definition
which expresses the individually necessary and jointly sufficient conditions for the word to apply. For instance, the definition of 'mother' could'
be female parent. If this is so, it can be represented by assignit;lg 'mother' as
the lexical entry for the concept female parent, or by associating with the
concept mother the elimination rule in (50) (where X and Y stand for
possibly empty strings of constituents):
(50) Mother-elimination rule
Input: (X - mother - Y)
Output: (X - female parent - Y)
On this classical view, proper names, for which necessary and sufficient
conditions of application cannot be given, are radically different from
other words: they have reference but no meaning. 'Homer', for instance,
cannot be defined as the author of the Iliad and the Odyssey, since there is
no inconsistency in denying that he is their author; he cannot be defined as
a Greek man (a grossly incomplete definition anyhow), since there is no
inconsistency in denying that he was Greek or that he was a man, and so
on. If that is so, if there are words which have reference but no logical

91
conditions attached, again, our approach can handle them easily by
associating them with concepts which have an empty logical entry.
Against the classical view, it has been argued, by Saul Kripke and Hilary
Putnam in particular,15 that proper names are far from being unique. Their
reference is fixed by an initial act of 'baptism' and maintained by a causal
chain which relates each of their particular uses to this initial act. The same
is true, they argue, at least for natural kind terms such as 'salt" or 'giraffe':
you learn the meaning of 'salt' by being shown typical samples of salt and
being told that they are called 'salt'; your teacher learned in the same way,
and all adequate uses of 'salt' are linked by a causal chain to some initial
'baptism'. True, a chemist might provide a definition of 'salt', but it is not
necessary to know this scientific definition in order to use the term 'salt'
properly. When this definition is known at all, it is better regarded as
belonging to the encyclopaedic entry of the concept, as knowledge of
what salt is, rather than as knowledge of what 'salt' means. If this 'causal
theory of reference' is entirely correct, then the meaning of natural kind
terms could be represented in our framework in terms of concepts with
empty logical entries and appropriate encyclopaedic entries.
According to various versions ot 'prototype theory', 16 again, the
meaning of a word is determined not by a set of logical properties, but by a
mental model of the thing the word is used to refer to. You have a mental
model of a prototypical giraffe, say, and you use the word 'giraffe' to refer
to things which resemble your model. If the meaning of some words is so
characterised, then, in our framework, they will have an empty logical
entry and an encyclopaedic entry containing the required model. 17
It seems to us, though, that both the causal theory and the prototype
theory might well tell only part of the story, and that even in their pet cases
- giraffes etc. -logical properties may still play a role. This makes us feel
closer to the view advocated by Fodor and his colleagues,18 that the
meaning of most words cannot be defined in terms of, or decomposed
into, more primitive concepts: 'mother', 'bachelor' and a few oftenquoted cases are exceptional rather than typical in this respect. Take
'yellow'; assume that it can be defined in terms of more primitive
concepts: one of the concepts would undoubtedly be colour; what would
be the other(s)? Take 'giraffe'; assume that a giraffe is by definition an
animal; make it even by definition a quadruped, if you want: how do you
then complete the definition? The best you can do, it seems, is to define
'yellow' as the colour yellow, 'giraffe' as animal (or quadruped) of the
giraffe species, but these are not proper definitions since they make use of
the very concepts they purport to define. The conclusion is that the
meaning of a word such as 'yellow', 'giraffe' or 'salt' is an irreducible
concept. Such concepts have logical properties, but these do not amount
92
Inference
to a definition of the concept. Fodor et al. propose representing these

logical properties in terms of meaning postulates; they can as easily be
represented in terms of elimination rules, e.g.
(51) Salt-elimination rule

Input: (X - salt - Y)
Output: (X - substance of a certain kind - Y)
(52) Giraffe-elimination rule
Input: (X - giraffe - Y)
Output: (X - animal of a certain species - Y)
(53) Yellow-elimination rule
Input: (X - yellow - Y)
Output: (X - colour of a certain hue - Y)
Our framework allows for empty logical entries, logical entries which
amount to a proper definition of the concept, and logical entries which fall
anywhere between these two extremes: that is, which provide some logical
specification of the concept without fully defining it. We assume that this
range of possibilities actually exists in the human mind; how exactly it is
exploited, to what extent actual concepts are logically specified, we see as a
matter for empirical research. What is at issue in the case of each concept
is: what deductive inferences are made possible by its presence in an
assumption ?19
In this section we have argued that the rules used in spontaneous
deductive inference are elimination rules attached to concepts. We have
treated concepts as triples of entries, logical, lexical and encylopaedic, filed
at an address. In one sense, the distinction between address and entry is a
distinction between form and content, the address being what actually
appears in logical forms, and the various entries spelling out its logical,
lexical and encyclopaedic content. In another sense, though, everything
discussed in this section has been purely formal. Logical entries are sets of
deductive rules: that is, formal operations on logical forms; encyclopaedic
entries are sets of assumptions: that is, representations with logical forms;
and lexical entries are representations with linguistic forms. All three
types of entry are thus available for use in a computational account of
comprehension.
Occasionally, an entry for a particular concept may be empty or
lacking. For example, a concept such as and, which has no extension, may
lack an encylopaedic entry. We saw that proper names and other concepts
could be seen as having an empty logical entry. Finally, there may be
concepts which have encyclopaedic and logical entries and play a role in
cognitive processes, but which are not lexicalised and which therefore
have an empty lexical entry. For example, it seems reasonable to assume
that corresponding to the general concept lexicalised as 'the military' or
93
'the armed forces', we have a particular concept of a soldier! sailor/airman,

which lacks a lexical entry.
Although the boundaries between logical and encyclopaedic entries are
not always easy to draw, we have tried to show that there are principled
differences between them, and that they may be expected to play different
roles in comprehension. In the next section we continue our account of
spontaneous inference by looking at the deductive process it~elf.
5 The deductive device

To the extent that deduction has been considered at all in the pragmatic
literature, it has been tacitly modelled on informal (natural) deduction
systems of the type familiar from introductory logic texts (e.g. Lemmon
1965, Thomason 1970, McCawley 1980). An informal deduction system
consists of a smallish set of deductive rules dealing with inferences which
hinge on the presence of such concepts as and, or, if . .. then, etc. Rules
(46)-(48) above (the rules of and-elimination, modus ponendo ponens and
modus tollendo ponens) are examples. Typically, no instructions are given
about how the rules are to be applied, in what order, or to what set of
assumptions as premises. Constructing a logical derivation in an informal
system is a matter of deciding what combination of rules and premises
might lead to interesting results. There is no way of predicting in advance
which premises will be chosen, which rules will be applied, and hence
which conclusions drawn.
Despite the widespread scepticism about the role of deductive reasoning in comprehension, many existing pragmatic theories, especially those
built on Gricean lines, seem to be based on informal systems of just this
type. When a certain inference or implicature is drawn, it can be shown ex
post facto how the hearer could have derived it from the premises available
at that point in the conversation by the use of available deductive rules.
However, it would almost invariably have been possible, from the same
set of premises, using the same set of rules, to derive quite different
conclusions, which would not in practice have been either intended or
drawn.
An informal system thus leaves an important part of the deductive
process unspecified: it is left to the intelligent user of the system to decide
how best to exploit it. In trying to construct a model of the mind, or the
part of the mind used in utterance comprehension, it is not legitimate to
rely on informal systems of this type, precisely because they leave an
important part of the comprehension process unexplained. Formal
systems (effective procedures, automata, algorithms) differ from informal
systems in just this respect: their procedures can be carried out by an
94
Inference
automaton whose decisions are predetermined at every stage. With a

formal system, it is decided in advance what assumptions are to be used as
premises; a set of assumptions are provided which, for the purposes of this
deduction at least, are to constitute the axioms or initial theses of the
system (hence such systems are often called axiomatic deduction systems).
It is also fully specified in advance which operations mayor must apply.
Nothing is left to the intuitions of the user: all the information necessary
for performing a deduction, all decisions involved in it, are fully specified
by the system itself.
The importance of formal systems for modell\ng mental abilities has
become increasingly apparent since Chomsky first used them in the study
of language. Chomsky insisted on the difference between informal,
traditional grammars and explicit, generative grammars. Informal grammars rely heavily on the intuitions of the user, and are intended to
supplement rather than account for these intuitions. They do not try to
make explicit what every speaker of a human language already knows, or
to rule out what no speaker of a human language would think of doing.
Just as informal deduction systems rely on the logical capabilities of the
user, so informal grammars presuppose a user with a considerable amount
of tacit linguistic knowledge which they make no attempt to explain.
Generative grammars, by contrast, are intended to give an explicit,
exhaustive account of the linguistic knowledge of the individual. A
generative grammar consists of a set of rules or principles designed to
provide a complete description of every sentence in a language, leaving
nothing about the structure of these sentences up to the intuitions of the
individual. Hence they are formal in the sense just described, and they
explain the individual's linguistic intuitions in a way that informal
grammars do not. It is not just that generative grammars, and more
generally formal systems, provide one way of modelling mental abilities.
There is no other way of modelling them known today; no other way of
accounting for intuitions than by providing a formal system that can itself
be operated without appeal to intuition.
What We want to offer here is the general outline of a formal deduction
system intended to model the system used by human beings in spontaneous inference, and in normal utterance comprehension in particular.
We are not proposing a fully described system but simply stating some of
its general properties. Nor, for that matter, has a fully described formal
grammar for a natural language ever been proposed: in both cases the
complexity of the phenomena to be described and the number of
theoretical choices to be made is enormous. What seems to us important at
this stage is to provide a general framework within which more detailed
hypotheses can eventually be produced and evaluated.
The device we envisage is an automaton with a memory and the ability
95
to read, write and erase logical forms, compare their formal properties,
store them in memory, and access the deductive rules contained in the
logical entries for concepts. Deductions proceed as follows. A set of \,.
assumptions 20 which will constitute the axioms, or initial theses, of the
deduction are placed in the memory of the device. It reads each of these
assumptions, accesses the logical entries of each of its constituent
concepts, applies any rule whose structural description is satisfied by that
assumption, and writes the resulting assumption down in its memory as a
derived thesis. Where a rule provides descriptions of two input assumptions, the device checks to see whether it has in memory an appropriate
pair of assumptions; if so, it writes the output assumption down in its
memory as a derived thesis. The process applies to all initial and derived
theses until no further deductions are possible.
The system monitors for redundancies and contradictions in its
derivations in the following way. Before writing an assumption down in
its memory, it checks to see whether that assumption or its negation is
already there. If the assumption itself is there, the device refrains from
writing it down again, and marks the theses and deductive rules used in
deriving it so that the derivation will not be repeated. If the negation of the
assumption is already there, the device halts, and the deductive process is
suspended until the ~ontradiction is resolved; a method of resolving
contradictions will be considered below. Subject to these constraints, the
device continues to operate until no new theses can be derived.
The move to formal systems raises questions about the capacity of the
deductive device which are sometimes overlooked when informal systems
are proposed. Most informal systems - at least those invented by logicians
- aim at completeness: that is, they aim to provide deductive rules which
will derive as logical implications all the entailments (or all those that hinge
on the logical properties of and, or, etc.) of a given set of assumptions. It is
easy to show that this set of entailments is infinite for any finite set of
premises. For example, a single arbitrary assumption P entails each of the
following conclusions:
(54) (a) (P and P)
(b) (P or Q)
(c) (not (not P))
(d) (If (not P) then Q)
(e) (If Q then P)
These are all entailments of P in the sense that there are no conceivable
states of affairs in which P would be true and any of (54a-e) false.
Logicians aiming at completeness will therefore set up deductive rules
enabling each of (54a-e) to be derived as logical implications of P. (54a) is
standardly derived by the rule of and-introduction, which takes two
96
Inference.
arbitrary assumptions, in this case P and P, as premises, and derives their
conjunction as conclusion:
(55) And-introduction
Input: (i) P
(ii) Q
Output: (P and Q)
(54b) is standardly derived by the rule of or-introduction, which takes an
arbitrary assumption as premise, and derives its disjunction with any
other arbitrary assumption as conclusion:
(56) Or-introduction
Input: P
Output: (P or Q)
(54c) is standardly derived by the rule of double negation, which takes an
arbitrary assumption as premise and derives the negation of its negation as
conclusion:
(57) Double Negation

Input: P
Output: (not (not P
Similar, though more complex, derivations yield (54d-e).
In informal systems, the existence of such rules creates no serious
problems, because it is left to the intelligent user to decide which line of
reasoning to pursue, and when to abandon it. However, in a formal system
of the type just described, the assumption is that although the rules may be
accessed and tested in a certain order, every rule applies obligatorily
whenever it is accessed and its input description is met. In such a system,
each of the above rules, once set in motion, would reapply indefinitely to
its own output, and the derivation would never stop.
Let us define an introduction rule as a rule whose output assumption
contains every concept contained in its input assumption(s), and at least
one further concept. We take the correct conclusion to be that introduction rules play no part in the spontaneous deductive processing of
information, the processing which our deductive device is designed to
describe. The only deductive rules available for use in the spontaneous
processing of information - the only rules which in any interesting sense
form part of the basic deductive equipment of humans - are elimination
rules.
This is a substantial claim, and one that most people working on the
psychology of deduction have been reluctant to make. If they consider the
problem at all, their solution is generally to make use of introduction
rules, but to constrain their functioning in some way so that indefinite

97
reapplication is avoided. We will discuss their reasons below. Note,
however, that our rejection of introduction rules is not based on the sole
desire to avoid indefinite reapplication. Our claim is that introduction
rules are never used in the spontaneous processing of information. For
example, no speaker would utter (58) expecting any of the conclusions in
(59a-e) to be drawn on the basis .of this utterance alone, and no hearer
would draw such conclusions on the basis of this utterance alone:
(58)The Prime Minister has resigned.
(59)(a) The Prime Minister has resigned and the Prime Minister has
resigned.
(b) Either the Prime Minister has resigned or it's a little warmer
today.
(c) It's not true that the Prime Minister hasn't resigned.
(d) If the Prime Minister hasn't resigned, the tiger will become
extinct.
(e) If it's the Queen's birthday, the Prime Minister has resigned.
The conclusions in (59a-e), and others derived by use of introduction
rules, are in some intuitive sense trivial. The intuition of triviality relates to
the fact that they leave the content of their input assumptions unchanged
except for the addition of arbitrary material; they can in no sense be
regarded as analysing or explicating the content of their input assumptions. Elimination rules, by contrast, are genuinely interpretive: the
output assumptions explicate or analyse the content of the input
assumptions. Our hypothesis is that the human deductive device has
access only to elimination rules, and yields only non-trivial conclusions,
defined as follows:
(60) Non-trivial logical implication
A set of assumptions P logically and non-trivially implies an
assumption Q if and only if, when P is the set of initial theses in a
derivation involving only elimination rules, Q belongs to the set of
final theses.
In other words, the human deductive device is a system which explicates
the content of any set of assumptions submitted to it.
Psychologists who have proposed models of the human deductive
system have rarely come to this conclusion. In fact, most of them are
concerned less with the study of spontaneous comprehension than with
performance on specific reasoning tasks: syllogistic reasoning, distinguishing valid from invalid arguments, and so on. The fact that these are
finite rather than open-ended tasks, and that many of the models proposed
are informal rather than formal, has meant that the problem of trivial
implication is often overlooked. Where it is not, the conclusion has almost
98
Inference
invariably been that introduction rules cannot be entirely dispensed with,

for two sorts of reasons. First, it is claimed that certain types of
spontaneous deduction which are regularly and straightforwardly performed require the use of introduction rules. For example, it is intuitively
clear that given premises (61 a-c), conclusion (62) would be spontaneously
derived, and that given premises (63a-b), conclusion (64) would be
spontaneously derived (subject to the usual limitations on memory and
attention) :
(61) (a) If the trains are on strike and the car has broken down, there is no
I
way of getting to work.
(b) The trains are on strike.
(c) The car has broken down.
(62) There is no way of getting to work.
(63) (a) If the boiler needs repairing or the electricity has been cut off, the
house will be uninhabitable.
(b) The boiler needs repairing.
(64) The house will be uninhabitable.
These examples have been seen as clearly demonstrating the need for at
least some version of the rules of and-introduction and or-introduction in
any account of the human deductive device.
The assumption is that the only way, or the only psychologically
plausible way, to derive (62) from (61) or (64) from (63) is by the use of
introduction rules. The derivation of (62) would involve a step of
and-introduction at (d) below, followed by a step of modus ponens based
on (a) and (d) to reach the desired conclusion:
(61) (a) If the trains are on strike and the car has broken down, there is no
way of getting to work. [Premise]
(b) The trains are on strike. [Premise]
(c) The car has broken down. [Premise]
(d) The trains are on strike and the car has broken down. [By andintroduction from (b) and (c)]
.
(62) There is no way of getting to work. [By modus ponens from (a) and
(d)]
Similarly, the derivation of (64) is seen as involving a step of orintroduction at (c) below, followed by a step of modus ponens based on (a)
and (c) to reach the desired conclusion:
99
(63) (a) If the boiler needs repairing or the electricity has been cut off, the
house will be uninhabitable. [Premise].
(b) The boiler needs repairing. [Premise]
(c) The boiler needs repairing or the electricity has been cut off. [By
or-introduction from (b)]
(64) The house will be uninhabitable. [By modus ponens from (a) and (c)]
These are, of course, the simplest derivations available in most standard

logics using primitive rules alone.
However, to show that the rules of and-introduction and orintroduction are necessary, it would first have to be shown that there is no
alternative derivation using only elimination rules, or that any such
alternative derivation was psychologically unmotivated. As regards the
first point, alternative derivations undoubtedly exist. Any standard logic
would permit the use of the following derived rules:
(65) Conjunctive modus ponens
(a) Input:
(i) (if (P and Q) then R)
(ii) P
Output: (If Q then R)
(b) Input:
(i) (If (P and Q) then R)
(ii) Q
Output: (If P then R)
(66)Disjunctive modus ponens
(a) Input:
(i) (If (P or Q) then R)
(ii) P
Output: R
(b) Input:
(i) (If (P or Q) then R)
(ii) Q
Output: R
These rules, like modus ponens itself, are elimination rules, and as will be
seen below, there is good reason to think that they play a role in the
spontaneous deductive processing of information. We assume that some
version of rule (65) is attached to the logical entry for and, and some
version of rule (66) is attached to the logical entry for or.
Rules (65) and (66) make it possible to derive (62) from (61) and (64)
from (63) without the use of introduction rules. The derivation of (62)
from (61) would go as in (61') below, with a step of conjunctive modus
ponens at (b'), followed by a step of regular modus ponens:
(61') (a) If the trains are on strike and the car has broken down, there is no
way of getting to work. [Premise]
100
Inference
(b) The trains are on strike. [Premise]
(b') If the car has broken down, there is no way of getting to work.
[From (a) and (b) by conjunctive modus ponens]
(c) The car has broken down. [Premise]
(62) There is no way of getting to work. [From (b') and (c) by modus
ponens]
The derivation of (64) from (63) would go as in (63') below, with a step of
disjunctive modus ponens deriving the conclusiqn directly from the
premIses:
(63') (a) If the boiler needs repairing or the electricity has been cut off, the
house will be uninhabitable. [Premise]
(b) The boiler needs repairing. [Premise]
(64) The house will be uninhabitable. [From (a) and (b) by disjunctive
modus ponens]
There is thus no question that alternative derivations exist.
The psychological plausibility of these derivations depends on the
psychological plausibility of rules (65) and (66) themselves. Rips (1983)
has experimental evidence that rule (66), the rule of disjunctive modus
ponens, is 'not only psychologically real but is one of the most highly
accessible rules, more accessible than the rule of modus ponens itself. His
evidence also shows that the rule of or-introduction is one of the least
accessible rules, and is indeed rejected by many subjects. The fact that
derivations such as (63)-(64) are regularly and easily performed strongly
suggests that no step of or-introduction is involved.
We know of no experimental evidence on rule (65), the rule of
conjunctive modus ponens. However, in a relevance-based framework,
both conjunctive and disjunctive modus ponens would be highly valued
for the following reason. When some item of information is presented in
the form of a complex conditional with a conjunctive or disjunctive
antecedent, the chances of finding the whole conjunctive or disjunctive
antecedent ready-stored in memory are clearly much smaller than those of
finding just one of its constituent conjuncts or disjuncts. What the rules of
conjunctive and disjunctive modus ponens do is allow inferences to be
drawn on the basis of a single conjunct or disjunct, rather than requiring
the whole conjunctive or disjunctive antecedent to be supplied. They thus
increase the chances of the presented information interacting with the
individual's existing representation of the world to enable new conclusions to be drawn. For an organism interested in improving its representa-

101
tion of the world, rules (65) and (66) would thus have considerable value.
The other justification for introduction rules found in the psychological
literature is based on subjects' performance on specific reasoning tasks: in
particular, in checking the validity of arguments. Thus, the fact that almost
all subjects say that (67a-b) entails (68), and some say that (69) entails (70),
is seen as evidence that the rules of and-introduction and or-introduction
are psychologically real: 21
(67) (a) Snow is white.
(b) Grass is green.
(68) Snow is white and grass is green.
(69) The world is round.
(70) The world is round or the world is flat.
Psychologists aware of the problem of trivial implication then assume that
the functioning of these rules is constrained in one way or another to avoid
iterative application. One solution is to allow the system to formulate
goals, and to allow introduction rules to operate only when (i) their input
descriptions are met and (ii) the system has the specific goal of deriving a
conjunctive or disjunctive conclusion, say, because it is checking the
validity of an argument such as (67)-(68) or (69)-(70). Rules which operate
under such constraints are called 'backwards' rules and distinguished from
the regular 'forwards' rules which simply apply whenever their input
descriptions are metY
We do not doubt that hearers sometimes want to derive a particular
conclusion from an utterance, and have to have some procedure for
obtaining it. What we do doubt is that such a procedure is likely to involve
entirely different deductive rules from those used in normal comprehension: that is, that a set of deductive rules exists whose sole function is to
confirm conclusions which they cannot spontaneously generate. It seems
much more likely that 'backwards reasoning' is merely the search for a set
of premises from which the desired conclusion can be derived using the
regular deductive rules: in other words that it is a retrieval strategy rather
than a distinct form of reasoning. And if, as in the cases (67)-(70) above,
there is no elimination rule which will directly derive the desired
conclusion from the set of available premises, then that conclusion is
simply not directly derivable.
However, there is no reason to think that the only method individuals
have of checking the validity of arguments is by direct derivation. An
argument is valid if and only if the premises entail the conclusion: that is, if
and only if the conclusion must be true whenever the premises are true.
We have argued that the human deductive device is incomplete in the sense
that there are valid arguments whose conclusions are not directly
102
Inference
derivable by the rules of the deductive device; (67)-(68) and (69)-(70) are
cases in point. We would therefore expect the deductive device to be
complemented with some non-deductive, or not directly deductive,
procedures for checking validity whenever the deductive machinery-is
insufficient. And in that case, from the fact that subjects make correct
judgements of validity, it does not follow that these judgements have been
arrived at by direct derivation.
OUr deductive device offers such an indirect procedure based on the fact
that it monitors contradictions. One way of showing that an argument is
valid is to show that it is inconsistent to assert the prquises while denying
the conclusion. For instance, if (67a-b) and the negation of (68), or (69)
and the negation of (70), were the initial theses of a derivation, the
deductive device would reveal the inconsistencies involved, and thus
establish that (67a-b) entails (68) and that (69) entails (70).
We therefore reject two extreme views of the human deductIve ability.
We do not believe that all deductive inference must be accounted for
purely in terms of deductive rules (the position tacitly adopted in Rips
1983). On the other hand, we do believe that a deductive rule system is an
extremely efficient device for reducing the number of assumptions that
have to be separately stored in memory, for accessing the conclusions of
arguments, for drawing out the implications of newly acquired conceptual
information, and for increasing the impact of this information on a stored
conceptual representation of the world. We therefore reject the claim
made by Johnson-Laird (1982b, 1983) that there are no mentally
represented deductive rules at all:
The crux of the matter is that a system of inference may perform in an
entirely logical way even though it does not employ rules of
inference, inferential schemata, meaning postulates, or any other
sort of machinery conventionally employed in a logical calculus.
Oohnson-Laird 1982b: 20)
It seems reasonable to assume, with Johnson-Laird, that subjects use
various heuristics that are not directly derivational in performing certain
types of reasoning tasks; but it does not follow that there are n~ mentally
represented deductive rules at all, any more than it follows from the fact
that subjects perform correctly on certain reasoning tasks that they must
be using deductive rules.
We are suggesting, then, a mixed view of human deductive abilities. Our
hypothesis is that when presented with a set of assumptions, subject to the
usual limitations of memory and attention,23 the device should directly and
automatically compute the full set of non-trivial implications defined by
its deductive rules, as part of its regular working procedure. Trivial
103
implications, by contrast, are not directly computed. The procedures for

accessing and checking the validity of trivial implications are quite
different from the automatic procedures for deriving the available
non-trivial ones: they are in a sense less natural, they may take more time
and be subject to different types of mistakes. In other words, performance
with the two types of implication should be significantly different, and
this might be experimentally checked.
In this section we have described in very general terms a deductive
device which might be used in the spontaneous processing of information.
The function of the device is essentially to analyse and manipulate the
conceptual content of assumptions, this function being performed by the
elimination rules attached to the logical entries for concepts. Our central
claim has been that in normal circumstances the deductive processing of an
assumption involves computation of its non-trivial implications, never of
its trivial ones. When an assumption is processed in a context of other
assumptions, again we claim that in normal circumstances only non-trivial
implications are computed. Since, if we are right, trivial implications play
no role in the comprehension process, we will not be concerned with them
in the rest of the book. From now on, unless otherwise stated, when we
talk of implications or logical implications, we will mean non-trivial
implications as defined above.
To restrict the cla:ss of implications that could in principle be computed
by the human deductive device is not, of course, to say everything about
the deductive processing of information. Quite apart from the need to
establish which deductive rules actually exist, there is the fact that the
implications of a given set of assumptions must be accessed in some order,
and we have as yet said nothing about how this order is imposed.
Moreover, information is invariably processed in a context of other
assumptions, and we have as yet said nothing about how the context is
selected. What we have done is merely place an upper bound on the set of
implications that could in principle be derived from a given set of
assumptions. How the premises are chosen, and in what order the
implications are computed, will be the subject of later chapters. In the next
section we want to consider, in rather more general terms, what types of
deduction can be performed when a chosen propositional content and
context are brought together in the memory of the deductive device.
6 Some types of deduction

We have claimed that the relevance of new information to an individual is
to be assessed in terms of the improvements it brings to his representation
of the world. A representation of the world is a stock of factual
Inference
assumptions with some internal organisation. We would now like to
suggest that the improvements brought by new information to an existing
representation of the world can be traced via the workings of the deductive
device.
When a set of assumptions is placed in the memory of the deductive
device, all the deductive rules in the logical entries attached to their
constituent concepts are accessed. As can be seen from the examples given
above, these rules are of two formally distinct types, which we will call
analytic and synthetic. An analytic rule takes only a single assumption as
input; a synthetic rule takes two separate assul\lptions as input. For
example, and-elimination (rules (46a-b) above), which takes a single,
conjoined assumption as input, is an analytic rule, and modus ponendo
ponens (rule (47) above), which takes a conditional assumption and its
antecedent as input, is a synthetic rule.
Let us say that any conclusion obtained from an initial set of
assumptions by a derivation in which only analytic rules are used is
analytically implied by that set of assumptions:
104
(71) Analytic implication

A set of assumptions P analytically implies an assumption Q if and
only if Q is one of the final theses in a deduction in which the initial
theses are P, and in which only analytic rules have applied.
Notice that by this definition every assumption analytically implies itself.
Then any implication which is not analytic is synthetic:
(72) Synthetic implication

A set of assumptions P synthetically implies an assumption Q if and
only if Q is one of the final theses in a deduction in which the initial
theses are P, and Q is not an analytic implication of P.
In practice, this means that a synthetic implication is the result of a

derivation in which at least one synthetic rule has applied.
Inferential approaches to comprehension have sometimes been charged
with a failure to distinguish understanding an assumption or an utterance
from grasping its logical consequences. Since we have denied the
distinction between 'logical' and 'non-logical' terms and conflated
'logical' with 'semantic' deductive rules, we might seem' particularly
vulnerable to this charge. In fact, the problem is already partly solved by
our distinction between trivial and non-trivial implications, and our claim
that only the latter are involved in the comprehension process. However,
there is still an intuition that even among non-trivial implications, some
are more intimately connected with understanding a set of assumptions
and others with working out its logical consequences. This intuition can
105
be explained in terms of our distinction between analytic and synthetic

implications.
"
The analytic implications of a set of assumptions are those that are
necessary and sufficient for understanding it, for grasping its content.
Someone who claims to understand an assumption but denies one of its
analytic implications cannot be said to have understood it at all. Failure to
grasp the synthetic implications of a set of assumptions, by contrast, is not
a failure to understand the information being offered, but a failure to
exploit it to the full.
The synthetic implications of a given set of assumptions are those whose
derivations involve the application of at least one synthetic rule. For
example, (73a-c) synthetically implies (74a-b), the synthetic rule in
question being modus ponens:
(73) (a) There's a bus coming.
(b) If there's a bus coming, we'll get to work on time.
(c) If we get to work on time, it won't matter that we overslept.
(74) (a) We'll get to work on time.
(b) It won't matter that we overslept.
Or, to give a 'non-logical' example, (75a-c) synthetically implies (76a-c)
the synthetic rule involved being something like the rule of containment
shown in (77):
(75) (a) The ticket is in the wallet.
(b) The wallet is in the suitcase.
(c) The suitcase is in the car.
(76) (a) The ticket is in the suitcase.
(b) The ticket is in the car.
(c) The wallet is in the car.
(77) Containment rule 24
Input: (i) (X - is - in - Y)
(ii) (Y - is - in - Z)
Output: (X - is - in - Z)
The ability to understand the set of assumptions (75a-c) certainly
involves the ability to grasp such analytic implications as (78 a-f) :
(78) (a)
(b)
, (c)
(d)
(e)
(f)
The ticket is somewhere.

Something is in the wallet.
The wallet is somewhere.
Something is in the suitcase.
The suitcase is somewhere.
Something is in the car.
106
Inference
Someone who accepted (75a-c) but denied any of (78a-f) would be guilty
of a failure not of logic but of understanding. Ori the other hand, someone
could quite well understand (75a-c) without having computed the
synthetic implications (76a-c). Suppose you acquired each of these
assumptions at a different time and in different circumstances, so that you
never happened to bring them together and compute. their synthetic
implications (76a-c). This omission would not mean that you had
understood each individual assumption any the less. We all have hundreds
of thousands of assumptions stored in memory, from which hundreds of
thousands of synthetic implications could be computfd if only they could
all be brought together in the memory of the deductive device. The fact
that they never have been, and indeed never will be, does not mean that
each individual assumption has not been properly understood.
Notice that what makes a synthetic implication synthetic is not the form
in which its premises are presented but the nature of the rules used in
deriving it. There is no reason why a single complex assumption should
not have synthetic implications. For example, the conjoined assumption
in (79) synthetically implies (76a), just as the separate assumptions (75a)
and (75b) do:
(79) The ticket is in the wallet and the wallet is in the suitcase.
(75) (a) The ticket is in the wallet.
(b) The wallet is in the suitcase.
(76) (a) The ticket is in the suitcase.
The only difference in the way this implication is derived frorp. (75a-b) on
the one hand and from (79) on the other is that (79) must undergo
and-elimination before rule (77) can apply. Otherwise the derivations are
identical, and what is a synthetic implication in one case remains a
synthetic implication in the other.
We have now reached the point where a single assumption can have
three types of logical implication: trivial implications, which are not
directly computed by our device; analytic implications, which are
necessary and sufficient for understanding it; and synthetic implications,
which have to do not so much with grasping the information being offered
as with exploiting this information to the full. Our framework thus sheds
some light on the rather hazy pretheoretical distinction between 'semantic' and 'logical' implication, between intrinsic meaning and wider import.
The distinction between analytic and synthetic implications has an
important practical consequence. The analytic implications of a given
assumption are intrinsic to it: they are recoverable as long as the
assumption itself is recoverable, simply by reprocessing it through the
deductive device. Synthetic implications, by contrast, are not intrinsic to
any single member of the set of assumptions from which they are derived
107
(barring the case of conjunctions such as (79)). A synthetic implication is

necessarily based on two distinct elementary assumptions, and deriving it
. is not just a matter of having these assumptions somewBere in memory:
they have to be brought together in the small working memory of the
deductive device. Once there, there is no guarantee that they will ever be
brought together again, and their synthetic implications may well be lost
for ever if not computed on the spot.
We assume, as do most current models of memory, that information is
broken down as far as possible into smaller units before being stored in
memory, so that a conjoined assumption, for example, is not stored as a
unit but is broken down into its constituent conjuncts, which may end up
in different encyclopaedic entries. Any organism interested in improving
its overall representation of the world must therefore be interested in
recovering as many synthetic implications as possible from any set of
assumptions it is currently processing, before the set is dismantled for
separate storage. Analytic implications, by contrast, are only worth
recovering as a means to an end, the end being the recovery of further
synthetic implications.
As we have seen, assumptions entering the memory of the deductive
device have four possible sources: they can come from perception,
linguistic decoding or encyclopaedic memory, or they can be added to the
memory of the device as a result of the deductive process itself. In an
intuitive sense, assumptions derived or retrieved from encyclopaedic
entries are old information, whereas assumptions derived from perception
or linguistic decoding, i.e. from input systems, are newly presented
information, and become old in the course of being processed. In this
book, we are concerned with the effect of newly presented information, in
particular of assumptions derived via the linguistic input system, on old
information drawn from an existing representation of the world.
We want to look, then, at the effect of deductions in which the set of
initial theses placed in the memory of the deductive device can be partitioned into two subsets, P and C, where P may be thought of as new
information, and C as old information. Let us call a deduction based on
the union of P and C as premises a contextualisation of P in the context
C. The contextualisation of P in C may yield new conclusions not derivable from either P or C alone. These we will call the contextual implications
ofP in C:
(80) Contextual implication
A set of assumptions P contextually implies an assumption Q in the
context C if and only if
(i) the Union of P and C non-trivially implies Q,
(ii) P does not non-trivially imply Q, and
108
Inference
(iii) C does not non-trivially imply Q.
A contextual implication is new information in the sense that it could not

have been derived from C, the stock of existing assumptions, alone;
however, it is not just new information, since it is neither an analytic nor
a synthetic implication of P, the newly presented information, alone. It is
a synthesis of old and new information, a result of interaction between
the two'.
From a logical point of view, the only slightly unusual feature of this
definition of contextual implication is that it partitions the premises of. a
synthetic implication into two distinct subsets, one rubs et being treated as
carrying the implication in the context of the other. Logically speaking, of
course, the two subsets are on a par: it is the union of P and C which
synthetically implies Q, and the distinction between the two subsets is a
pragmatic rather than a logical one. Newly presented information is seen
as adding to, and interacting with, information drawn from an existing
representation of the world.
The addition of new information to a context of old information brings
not only contextual implications but also analytic, and perhaps synthetic,
implications of its own. However, these implications, unlike contextual
implications, are context-independent. It is mainly in terms of contextual
implications that the effect of context on utterance interpretation, and the
rationale for processing information in one context rather than another,
must be seen. The notion of contextual implication will play a major role
in the remainder of this book.
A central function of the deductive device is thus to derive, spontaneously, automatically and unconsciously, the contextual implications
of any newly presented information in a context of old information. Other
things being equal, the more contextual implications it yields, the more
this new information will improve the individual's existing representation
of the world. 25
The deductive device is also at the centre of spontaneous nondemonstrative inference: it is a major source of assumptions, and its
processes affect the strength of both the initial and final theses of the
deductions it performs. It is to these aspects of inference that we now turn.
7 Contextual effects: the role of deduction
in non-demonstrative inference
Recall that a deduction based on the union of new information P and old
information C is a contextualisation of P in C. Such a contextualisation may give rise to what we will call contextuai effects. In this section we
Contextual effects
109
will introduce the notion of a contextual effect. In chapter 3 we will argue

that relevance can be characterised in terms of contextual effects.
The intuitive idea behind the notion of a contex~ual effect is the
following. To modify and improve a context is to have some effect on that
context - but not just any modification will do. As we have seen, the
addition of new information which merely duplicates old information
does not count as an improvement; nor does the addition of new
information which is entirely unrelated to old information. The sort of
effect we are interested in is a result of interaction between new and old
information. One such effect has already been described. Contextual
implications are contextual effects: they result from a crucial interaction
between new and old information as premises in a synthetic implication.
Intuitively, in the sort of framework we have been describing, there
should be two more types of contextual effect. On the one hand, new
information may provide further evidence for, and therefore strengthen,
old assumptions; or it may provide evidence against, and perhaps lead to
the abandonment of, old assumptions. In the last three sections, we have
largely ignored the fact that assumptions placed in the memory of the
deductive device come with varying degrees of strength, and that a
deduction may result in,a ~ontradiction. We now want to look at these two
aspects of deduction and the type of contextual effects they give rise to.
We will then propose a gerieral characterisation of the notion of a
contextual effect.
How does the relative strength of the premises in a deduction affect the
strength of the conclusions? This question can be approached from either
a logical or a cognitive point of view. A better understanding of the logical
issue should help with the cognitive one.
Consider the set of premises (81a-d) and their synthetic implication
(82):
(81) (a) If Peter, Paul and Mary came to the party, the party was a success.
(b) Peter came to the party.
(c) Paul came to the party.
(d) Mary came to the party.
(82) The party was a success.
Con,sider (83), which is the conjunction of (81a-d), and which therefore
also entails (82):
(83) If Peter, Paul and Mary came to the party, the party was a success, and
Peter, Paul and Mary came to the party.
If we could assign a confirmation value to (83), it would be easy to show

that, from a logical point of view, the confirmation value of (82) must be at
110
Inference
least as high as that of (83), which entails it. It could not be lower, since this
would mean that conceivably (83) might be true and (82) false, and this is
ruled out by the fact that (83) entails (82). On the other hand, if there is any
possibility that the party was a success without Peter, or Paul, or Mary
being there, then the confirmation value of (82) should be even higher than
that of (83).
From a logical point of view, then, there is a lower limit to the
confirmation value of a conclusion: it cannot be less than the confirmation
value of the conjunction of the premises. From a cognitive point of view,
the question now becomes: how can the deductive de.vice assess that lower
limit, given that (as we are assuming) it can neither derive the conjunction
of the premises nor compute its confirmation value? Further logical
considerations must be taken into account before this cognitive question
can be answered.
We are considering, then, how the confirmation value of the conjunction (83) might be assessed. The confirmation value of a conjunction of
assumptions depends on the values of its conjuncts. From a logical point
of view, it cannot be higher than the value of the weakest, i.e. least
confirmed, conjunct. Suppose it is certain that Peter and Paul came to the
party, but doubtful that Mary did. It is equally doubtful, then, that Peter,
Paul and Mary all came to the party. On the other hand, the confirmation
value of a conjunction can be lower than that of its weakest conjunct.
Suppose that (81 b-d) are are all strongly confirmed, but less than certain.
Generally, the likelihood of all three assumptions being true is less than
the likelihood of any single one of them being true. The confirmation
value of a conjunction should therefore be lower than thatof any of its
individual conjuncts. In fact, the greater the number of conjuncts, and the
lower their confirmation values, the lower the confirmation value of the
conjunction.
There is thus an upper limit to the confirmation value of the conjunction
of premises used in a deduction, and this can be assessed without either
deriving that conjunction or computing its confirmation value. The upper
limit of the confirmation value of a conjunction is the confirmation value
of its weakest conjunct; to assess that, no computation is required.
Notice that in a single deduction, different conclusions may be derived
on the basis of different premises. Only those premises actually used in
the derivation of a particular conclusion should affect its confirmation
value. For instance, let us add to (81) one further premise (81e):
(81) (e) If Paul and Mary came to the party, Roger left early.
Now, from the set of premises (81a-e), another conclusion follows:
(84) Roger left early.
Contextual effects
111
Clearly, we do not want the confirmation value of (84) to be affected by

that of (81a) or (81b), nor do we want the confirmation value of (82) to be
affected by that of (81e). Thus, a particular conclusioncoshould inherit its
confirmation value only from the set of premises actually used in its
derivation.
All this suggests a cognitive account, based on the working of the
deductive device, of the relation between the strength of the premises and
the strength of the conclusions in a deduction. Such an account could be
implemented along the following lines. The deductive device might
operate in such a way that when an analytic rule is applied, the conclusion
inherits the strength of the premise. When a synthetic rule applies, there
are three possibilities. Either both premises are certain, in which case the
conclusion is also certain; or one of the premises is certain and the other is
not, in which case the conclusion inherits the strength of the weaker
premise; or neither premise is certain, in which case the strength inherited
by the conclusion is lower than that of the weaker premise.
The effect on derivations involving the application of more than one
rule would be as follows. When all the premises actually used in the
derivation of a particular conclusion are certain, the conclusion is also
certain. When all the premises but one are certain, the conclusion inherits
the strength of the less-than-certain premise. When more than one
premise is less than certain, then the conclusion is weaker than the weakest
premise. Conclusions derived from several weak premises inherit a value
that is very weak and vague. However, inherited degrees of strength are
lower limits: generally speaking, conclusions are more likely to be true
than the conjunction of the premises from which they are deduced.
We are crediting the deductive device with the ability to reproduce
degrees of strength, and to raise or lower them. These are much more
modest abilities than would be needed for computing quantitative
confirmation values. As expected, these procedures do not determine
absolute confirmation values except of the grossest kind (e.g. weakly
confirmed, certain), but by pushing the strength of a given assumption
above or below that of other assumptions, it makes some comparisons
possible.
When a conclusion is derived from severalless-than-certain premises,
its value will be very vague. This would be a great defect if we were trying
to develop an optimal logical system. But we are actually trying to model a
cognitive system. The inability of our model to determine precisely the
strength of a conclusion derived from several uncertain assumptions
corresponds well enough to introspective evidence. We are not aware of
any other evidence showing that the human mind is equipped with a more
powerful and precise way of spontaneously determining the strength of its
factual assumptions.
112
Inference
Returning to our discussion of contextual effects, consider first how

contextual implication, our original example of a contextual effect, fares
when the strength of assumptions is taken into account. Contextual
implication is a sub-type of synthetic implication. Consider, then, the
relation between a synthetic implication and the premises actually used
in deriving it. A synthetic implication is not logically implied by anyone
of its premises: it cannot be demonstratively inferred from anyone of its
premises. On the other hand, take away anyone of the premises, and the
conclusion can no longer be derived from the remaining premises. It could
thus be said that each premise is an argument. for the conclusion, or
evidence for the conclusion, in the context of the other premises; or, as we
propose to say, each of the premises actually used in the derivation of a
synthetic implication strengthens the conclusion which they jointly
imply. The contribution of each individual premise to the strength of this
joint conclusion is a function of its own degree of strength.
Contextual implication, as we have defined it, is a relation between a
synthetic implication and one of the premises used in deriying it. A
contextual implication Q cannot be demonstratively inferred from an
assumption P that contextually implies it in context C; but it can
be demonstratively inferred from the union of P and C; or it can be nondemonstratively inferred from P, by assuming C. Here, then, is a
case of non-demonstrative inference in which the only logical rules
involved are deductive. These rules contribute not only to the formation
of a new assumption on the basis of existing assumptions, but also to the
determination of its strength.
We see the relation of contextual implication, then, as a special case of
contextual strengthening. It might be called dependent strengthening, in
the sense that the strength of the conclusion depends not only on the
added premises P but also on the context C: P affects, but does not
fully determine, the strength of its contextual implication. Dependent
strengthening contrasts, of course, with the better known case of
independent strengthening - the cognitive counterpart of independent
confirmation - to which we now turn.
Independent strengthening arises when a single conclusion is independently implied by two different sets of premises. Consider the set of
premises (85a-b):
(85) (a) If the party broke up late, then it was a success.
(b) The party broke up late.
This logically implies (82), a conclusion which, as ",e saw, is also implied
by the set of premises (81a-d):
(81) (a) If Peter, Paul and Mary came to the party, the party was a success.
Contextual effects
113
(b) Peter came to the party.

(c) Paul came to the party.
(d) Mary came to the party.
Suppose, now, that (85a-b) is contextualised in (81a-d): will that
contextualisation have any effect?
We said that the deductive device has the following procedure for
avoiding redundancies: before writing an assumption down in its
memory, it checks whether that assumption is already there. If so, it
refrains from writing the assumption down again, and marks the theses
and deductive rules used in deriving it so that the derivation will not be
repeated. However, this discussion of what happens when the device
meets the same assumption twice did not take into account the possibility
that the two occurrences of this assumption may differ in strength. It is in
the handling of such cases that significant effects might occur.
The question, then, is how the strength of an assumption already
present in the memory of the deductive device, or deducible from the
theses present in it, should be affected when the device meets a second
occurrence of the same assumption, derived from different premises. Let
us call the strength that (82) inherits from (81a-d) alone S1, the strength
that it inherits from (85a-b) S2, and the strength it inherits from the union
of (81 a-d) and (85a-b) S3. The question then is how S3 is related to S1 and
S2.
Both intuitively and logically speaking, S3 should be greater than both
S1 and S2 (unless, of course, S1 or S2 is certain, in which case S3 should be
certain too). The reason for this is straightforward. First, whichever of S1
and S2 is the greater provides the lower limit for S3: if S3 fell below this
limit, it would fail to reflect the amount of support that each set of
premises independently brings to their common conclusion. Second, if S3
were merely identical to whichever of S1 and S2 were the greater, that is, if
it merely reflected the amount of support that one set of premises brought
to their common conclusion, it would entirely fail to reflect the support
that the other set of premises independently brings to it. Hence S3 should
be greater than both S1 and S2. In other words, (82) should inherit from
the union of (81a-d) and (85a-b) a degree of strength greater than the one
it inherits from either (81a-d) or (85a-b) independently. There is no
difbculty in implementing this logical condition through the working of
the deductive device as we have described it.
Dependent and independent strengthening can combine. Suppose that
(86a-?) is contextualised in (81 a-d), and that (81d) is the weakest of all the
premIses:
114
Inference
(86) (a) Either Bob came to the party or Mary came to the party.
(b) Bob did not come to the party.
(86a-b) logically implies (87):
(87) Mary came to the party.
This leads to an independent strengthening of (81d), which is identical in
content to (87). Since (81d) is also an argument for (82) in the context of
(81a-c), (82) is in turn dependently strengthened by (81d). The contextualisation of (86) in (81) thus strengthens both ~81d) and (82).
We have so far considered two types of contextual effect: the addition of
contextual implications and the strengthening of previously held assumptions. But a significant improvement of one's representation of the world
can also result from the elimination of false assumptions. This highly
significant contextual effect may be brought about when there is a
contradiction between new and old information.
In our account of the working of the deductive device, we said that
when it encounters a contradiction, it halts until the contradiction is
resolved. Suppose, for example, that (88a-b) is contextualised in (89):
(88) (a) If Jennifer came, the party was a success.
(b) Jennifer came.
(89) (a) If Bill came, the party was not a success.
(b) Bill came.
(c) The party was not a success.
(d) If the party was not a success, we won't have another party.
(e) We won't have another party.
(88a-b) logically implies (90), the negation of (89c):
Ort deriving (90), we said, the device will attempt to resolve this
contradiction. In resolving a contradiction, the strength of the two
contradictory assumptions must be taken into account.
The deductive device has the power not only to read and write
assumptions in its memory, but also to erase them. Let us assume that
when two assumptions are found to contradict each other, if it is possible
to compare their strengths, and if one is found to be stronger than the
other, then the device automatically erases the weaker assumption. When
an assumption is erased, the device also erases any assumption which
analytically implies it, and the weaker of any pair of assumptions which
synthetically imply it; this procedure applies recursively until no more
erasures can take place. When such a procedure is possible, the contradiction is eliminated at the root, and the deductive process can be resumed.
Contextual effects
115
Suppose, for example, that (90) is stronger than (89c). On discovering

the contradiction, the device would erase (89c). It would then check to see
whether it has in memory any assumption which analytically implies
(89c), or any pair of assumptions which synthetically imply it; it would
discover that (89c) is synthetically implied by (89a) arid (89b), and would
erase the weaker of these two assumptions.
Notice now that (89c) and (89d) synthetically imply (8ge): that is, (89c)
is an argument for (8ge) in the context of assumption (89d). What happens
to (8ge) when (89c) is erased? Clearly, (8ge) should lose whatever strength
it gained by dependent strengthening from (89c). It may be independently
strengthened by other assumptions; in which case it should remain in the
memory of the device with a degree of strength commensurate with that
independent support. Or it may have no other source of strength, having
received all its support from (89c) and (89d); in which case the loss of this
support should reduce it from the status of an assumption with some
degree of strength to the status of a mere unsupported possibility.
There are situations where this straightforward method of resolving
contradictions yields no result: for instance because the device is unable to
compare the strength of the two contradictory assumptions, or because
they are equally strong. We assume that in these situations the contradiction is resolved by other means: for example, by a conscious search for
further evidence for or against one of the contradictory assumptions. This
seems to correspond to the introspective evidence that some contradictions are resolved by an apparently immediate and automatic rejection of
the faulty premises, while other contradictions require deliberation.
The contextualisation of a new assumption in a context which
contradicts it can result in the rejection, not of an assumption already
present in the context, but of some or all of the new information itself. In
this case, there will be no significant contextual effect. Contextual effects
are achieved only when, as in the case just described, the new assumption
displaces an assumption already present in the context, with subsequent
weakening or erasure of other contextual assumptions linked to it by
relations of analytic or synthetic implication. If (90) had been weaker than
(89c), for example, (90) itself would have been erased, and the contextualisation of (88a-b) in (89a-e) would have had no effect at all.
We have now described the various types of possible contextual effects:
contextual implications, strengthenings, and contradictions resulting in
the erasure of premises from the context. cl We have so far considered two
types of strengthening: dependent and independent, in both of which the
strength of a conclusion is determined by the strength of the premises used
in deriving it. Before leaving the subject, we would like to suggest that
there is another type of strengthening, which we will call retroactive
strengthening. Here, the assumptions actually used in a contextualisation
116
Inference
may be strengthened by the fact that the contextualisation has had

some expected result. Anticipating the next chapters, we will briefly
outline how such retroactive strengthening can occur in verbal comprehension, and consider whether a similar effect could be occurring in
spontaneous inference generally.
In verbal communication, the hearer is generally led to accept an
assumption as true or probably true on the basis of a guarantee given by
the speaker. Part of the hearer's task is to find out which assumptions the
speaker is guaranteeing as true. Our hypothesis is that the hearer is guided
by considerations of relevance in carrying 0l.\t this task. He expects the
information the speaker intended to convey, when processed in the
context the speaker expected it to be contextualised in, to be relevant: that
is, to have a substantial contextual effect, at a low processing cost. Thus, if
the hearer assumes (91),
(91) The speaker intends to assert P
and P turns out to be relevant in the expected way, assumption (91) is
strengthened; moreover, if the hearer trusts the speaker to be truthful,
assumption P is strengthened too. If P turns to be relevant in the expected
way only when assumption Q is added to the context, then assumption
(92) is strengthened:
(92) The speaker intends the hearer to assume Q
and again, if the hearer trusts the speaker, then assumption Q is
strengthened.
What makes these retroactive strengthenings generally valid is the
following. It is generally unlikely that any arbitrary assumption would be
relevant enough to be worth the individual's attention; hence any
inrerpretation of an utterance that achieves a satisfactory level of relevance
is very likely to be correct. In other words, the hearer who arrives at an
adequately relevant interpretation can be fairly confident that it is the one
the speaker intended to convey. This point will be discussed in more detail
in chapter 4.
Now it might seem that the case of verbal understanding is unique and
significantly different from other uses of inferential abilities. The speaker
wants to be understood and is actively helping the hearer, whereas, so the
argument goes, the environment is not helping observers understand it. In
fact this argument is not as compelling as it looks. Much of the
environment is man-made and is full of intentional cues to help people
perceive it adequately. You might grant that and still ask, what abQut the
natural environment? Surely nature is not helping humans understand it?
Well again, that is not so sure.
Human cognitive abilities are a part of nature; they are well adapted as a
Contextual effects
117
result of natural evolution. It could be that of the assumptions which come

most spontaneously to a human mind, those that are true are more likely
to be relevant than those that are false, so that when relevance is achieved it
provides generally valid retroactive strengthening. If that is so, then
Fodor's suggestion that scientific thinking can be taken as typical of
central thought processes is dead wrong. Nature Itelps humans develop a
genuine but limited understanding of it - perfectly suited, say, to
palaeolithic hunters and gatherers. Science is an attempt to understand
nature more thoroughly but without nature's help, and hence without the
benefit of automatic retroactive strengthening.
Discussions by logicians of hypothesis formation and confirmation
have generally been inspired by the form these processes seem to take in
science. Yet scientific thinking may be quite different in relevant respects
from ordinary conceptual thinking. At least we would argue that it is
different from verbal understanding. In verbal understanding, nondemonstrative inference can be described without invoking any logical
rules apart from deductive rules; the strength of an assumption is a
by-product of the way it is formed and used, a by-product, in particular,
of the way it is deductively processed.
Having surveyed the various types of possible contextual effects, we are
now in a position to generalise. If all a contextualisation does is add all,
some or none of the new information to the context without otherwise
altering the context at all, then this contextualisation has no contextual
effect. Otherwise, there is some contextual effect, in the form of an erasure
of some assumptions from the context, a modification of the strength of
some assumptions in the context, or the derivation of contextual
implications. 26
In this chapter, we have presented a broad outline of the inferential
abilities which we assume are involved in spontaneous inference and, in
particular, in verbal understanding. We are very aware of its sketchiness,
and of the many questions it raises and leaves unanswered. It seems to us,
though, that these questions are not intractable, and that, on the bases we
suggest, the psychological study of non-demonstrative inference is an
interesting problem rather than an unfathomable mystery. We have also
characterised a notion of contextual effect around which, in the next
chapter, we will build an explicit notion of relevance.
3
Relevance
1 Conditions for relevance

In the last chapter, we introduced the notion of a contextual effect and
discussed a variety of such effects: contextual implications, contradictions, and strengthenings. The notion of a contextual effect is essential to a
description of the comprehension process. As a discourse proceeds, the
hearer retrieves or constructs and then processes a number of assumptions. These form a gradually changing background against which new
information is processed. Interpreting an utterance involves more than
merely identifying the assumption explicitly expressed: it crucially
involves working out the consequences of adding this assumption to a set
of assumptions that have themselves already been processed. In other
words, it involves seeing the contextual effects of this assumption in a
context determined, at least in part, by earlier acts of comprehension.
At each point in a discourse, the hearer has in the forefront of his
attention a different set of assumptions, which he may never have
processed together before, and may never process together again. By
working out the synthetic implications of this set of assumptions, he can
acquire new information which may be lost forever when that particular
set is dismantled and its constituent assumptions are either forgotten or
stored in their separate locations in the hearer's encyclopaedic memory.
It is not just that these assumptions come together in the hearer's mind
for what may be the only time. They also come together in a certain
sequence, and are presumably processed in that sequence, so that each new!'
assumption is processed in the context of a set of assumptions many of '
which have themselves just been processed. The notion of a contextual
effect helps describe these two essential properties of utterance comprehension: comprehension involves the joint processing of a set of
assumptions, and in that set some assumptions stand out as newly
Conditions for relevance
119
presented information being processed in the context of information that

has itself been previously processed.
The notion of a contextual effect is essential to a characterisation of
relevance. We want to argue that having contextual effects is a necessary
condition for relevance, and that other things beil1g equal, the greater the
contextual effects, the greater the relevance.
Before embarking on this project, we would like to make clear what we
are trying, and what we are not trying to do. We are not trying to define
the ordinary English word 'relevance'. 'Relevance' is a fuzzy term, used
differently by different people, or by the same people at different times. It
does not have a translation in every human language. There is no reason to
think that a proper semantic analysis of the English word 'relevance'
would also characterise a concept of scientific psychology.
We do believe, though, that scientific psychology needs a concept
which is close enough to the ordinary language notion of relevance; in
other words, we believe that there is an important psychological property
- a property of mental processes - which the ordinary notion of relevance
roughly approximates, and which it is therefore appropriate to call
relevance too, using the term now in a technical sense. What we are trying
to do is to describe this property: that is, to define relevance as a useful
theoretical concept.
We assume that people have intuitions of relevance: that they can
consistently distinguish relevant from irrelevant information, or in some
cases, more relevant from less relevant information. However, these
intuitions are not very easy to elicit or use as evidence. The fact that there is
an ordinary language notion of relevance with a fuzzy and variable
meaning is a hindrance rather than a help. Moreover, intuitions of
relevance are relative to contexts, and there is no way of controlling
exactly which context someone will have in mind at a given moment.
Asking people to restrict themselves to explicit, artificially constructed
contexts goes so much against natural procedures for context construction
that the resulting intuitions are of questionable value.
Despite these difficulties, we intend to invoke intuitions of relevance.
We should make clear, first, that when we claim that one assumption is
intuitively relevant and another not, or that one assumption is more
relevant than another, we merely expect you to perceive some difference;
whether you would ordinarily use the word 'relevance' to describe it is
beside the point. Second, we see these intuitive judgements of relevance as
suggestive and worth paying attention to, but we do not regard them as
conclusive. They will provide us with a starting point, but are certainly not
to be treated as a unique and final criterion. The value of our theoretical
notion of relevance will ultimately depend on the value of the psycholo-
120
Relevance
gical models which make use of it, and, in particular, on the value of the
theory of verbal comprehension that it allows us to formulate. Intuitions
o.f relevance are not the only kinds of intuition involved in comprehenSIOn.
If you take a particular set of assumptions C and add to it some arbitrarily

chosen assumption P, there is little reason to expect P to be relevant at all
in the context C, or to have any contextual effect in it either. For instance,
take C to be the set of assumptions you have in mind while reading this
sentence. Suppose we were now to tell you,
(1) 5 May 1881 was a sunny day in Kabul.
The assumption explicitly expressed by (1) is not likely to have any
contextual effect in C, or to be relevant (in any sense) in C. It is
intuitively obvious that the assumption expressed by (1) is irrelevant in
C. We can account for this by pointing out that (1) has no contextual
effect in c: there is no assumption in the context with which (1) might
combine to yield contextual implications; nor does it affect the'strength of
any assumption already present in the context. This is because (1) is utterly
unrelated to the context in question.
There are other ways in which an assumption may lack contextual
effects. Suppose we were now to tell you,
(2) You are now reading a book.
The assumption explicitly expressed by (2) is likely to be irrelevant in the

context of whatever assumptions you had in mind immediately before
reading it; this can again be accounted for by pointing out that it has no
contextual effect in that context. You were presumably already aware of
the fact that you were reading a book, so that any implications which (2)
would have had in that context would already have been computed.
Moreover, you presumably held this assumption as certain, so that its
strength could not be increased.
To take a third example, which is irrelevant for different reasons still,
suppose we were now to tell you,
(3) You are fast asleep.
The assumption explicitly expressed by (3) is inconsistent with a number

of unshakable assumptions which you currently have in mind. You are
presumably aware not only of the fact that you are now reading a book,
but also of the fact that this is an activity which is incompatible with being
fast asleep. Since, however much you trust us, on this question you would
rightly trust yourself more, the contradiction which results when the
assumption expressed by (3) is added to the present context would lead to
the erasure of (3), as described in the last chapter. In other words, (3)
Conditions for relevance
121
would have no contextual effect in the present context, and this is why it is
intuitively felt to be irrelevant.
'
There are thus t~!'~j~es of case in which. an assumption may lack
contextual effects, and be irrelevant, in a context>In the first, illustrated by
(1), the assumption may contribute new information; but this information
does not connect up with any information present in the context. In the
second, illustrated by (2), the assumption is already present in the context
and its strength is unaffected by the newly presented information; this
newly presented information is therefore entirely uninformative and, a
fortiori, irrelevant. In the third type of case, illustrated by (3), the
assumption is inconsistent with the context and is too weak to upset it;
processing the assumption thus leaves the context unchanged.
It should be stressed that in all these examples it is only the assumption
explicitly expressed by the utterance that lacks contextual effects and is
irrelevant: the fact that someone chooses to express an irrelevant
assumption may itself be highly relevant. For instance, it may be a way of
making manifest a desire to change the subject, and this desire may well be
relevant. Or, to take an actual example, we have expressed the irrelevant
assumptions (1)-(3) in an attempt to make what we hope were relevant
remarks. Relevance may be achieved by expressing irrelevant assumptions, as long as this expressive behaviour is itself relevant.
On the basis of these examples, we want to claim that an assumption
which has no contextual effect in a given context is irrelevant in that
context. In other words, having some contextual effect in a context is a
necessary condition for relevance.
.
The next question seems to be whether having contextual effects might
be not only a necessary condition for relevance but also sufficient. There is
a certain amount of evidence that it is. For example, consider the following
(attested) exchange:
(4) Flag-seller: Would you like to buy a flag for the Royal National
Lifeboat Institution?
Passer-by: No thanks, I always spend my holidays with my sister in
Birmingham.
To see the relevance of the passer-by's response, the hearer must be able to
supply something like the premises in (5), and derive something like the
contextual implication in (6):
(5) (a)
(b)
(c)
(d)
Birmingham is inland.
The Royal National Lifeboat Institution is a charity.
Buying a flag is one way of subscribing to a charity.
Someone who spends his holidays inland has no need of the
services of the Royal National Lifeboat Institution.
122
Relevance
(e) Someone who has no need of the services of a charity cannot be

expected to subscribe to that charity.
(6) The passer-by cannot be expected to subscribe to the Royal National
Lifeboat Institution.
What is interesting about the passer-by's reply is the very close connection
that exists between seeing its relevance (or, more precisely, the relevance
its speaker intended it to have) and being able to derive some contextual
implication from it. It seems clear that someone who is unable to supply
something like the context in (5) and derive something like the contextual
implication in (6) will be unable to see the inteAded relevance of this reply,
and that, conversely, anyone who sees this implication will agree that this
reply is relevant in the appropriate context. Perceiving some contextual
effect of an assumption seems to be sufficient for judging it relevant.
It might be tempting, then, to propose the following definition:
(7) Relevance
An assumption is relevant in a context if and only if it has some
contextual effect in that context.
This definition captures the intuition that to be relevant in a context, an
assumption must connect up with that context in some way. It clarifies
this intuition by specifying the nature of the connection required. For
example, it predicts that the passer-by's reply in (4) is relevant in the
context (5a-e) because it connects up with this context to yield the
contextual implication (6). In real life, of course, (6) would in turn be
processed in a context in which it would have further contextual
implications and other contextual effects: for example, strengthening or
weakening various assumptions of the hearer, thus ensuring the relevance
of the reply in a wider context.
Although the definition in (7) accords with some intuitive judgements
of relevance, we would expect there to be others which seem to go against
it, and particularly against the claim that having any contextual effect,
however small, is a sufficient condition for relevance. Intuitions about the
proper use of 'relevance' are like intuitions about the proper use of, say,
'flexibility': the more difficult it is to bend some object, the less willing we
are to call it flexible, even though we might recognise that, if an object can
be bent at all, then technically it is flexible. Intuitions about 'relevance' go
the same way: the weaker the contextual effects of anassumption, the less
willing we are to call it relevant, even though it can be argued that, if an
assumption has any contextual effect at all, then technically it is relevant.
Suppose, for instance, that we were to inform you now of (8):
(8) It took us a long time to write this book.
Degrees of relevance
123
At first sight, the assumption expressed by (8) seems irrelevant in the

context of the thoughts you have in mind if you are attending to the
present discussion of the notion of relevance. That it took us a long time to
write this book is not the kind of consideration you would expect at this
point.
However, we hope you share our intuition that (8) is not quite as
irrelevant as (1)-(3) (and if you do not, rest assured that nothing essential
hinges on it and just skip the rest of this paragraph). This can be related to
the fact that (8), unlike (1 )-(3), has some contextual effect in a context that
you are likely to have accessible. For instance, you might already have
suspected that it took us a long time to write this book. In that case, we are
independently strengthening this suspicion of yours, as well as any
implication you might already have derived from it. Or this might be
brand new information for you, in which case you might combine it with
whatever opinions you have already formed about the book and derive
some contextual implications: say, if you do not like the book, the
implication that we have been wasting a lot of time. Some contextual
effect, some relevance.
There are reasons, though, for trying to go beyond definition (7) which
are more compelling than debatable intuitions about the relevance or
irrelevance of (8). The intuitions of relevance that it is essential to account
for are intuitions not about the simple presence or absence of relevance,
but about degrees of relevance. It is to these that we now turn.
2 Degrees of relevance: effect and effort

The definition of relevance just proposed is insufficient for at least two
reasons: fi~t, beca~lev~IlCJ~j.s a matter~ ofd~gree, and we have said
nothing about how degrees of relevance are determined; second, because
we have defined relevance as a relation between an assumption and a
context, b~t have said nothing about how the context is determined. At
the moment, then, we have simply defined a formal property, whose
relation to psychological reality has been left undescribed.
Consider first the question of degrees of relevance. At a very general
level, we want to compare the concept of relevance to concepts such as
productivity or yield, which involve some form of cost-benefit analysis. A
firm with output of any value, however small, is productive to some
degree, just as we have claimed that an assumption with any contextual
effects at all, however limited, is relevant to some degree. However, where
the output is very small, there is some initial reluctance to say that the firm
is productive at all, even though, when compared to a firm with genuinely
124
Relevance
zero output, it is obviously productive to some degree: the parallel with

relevance is clear.
The output of the firm, i.e. the value of the goods produced, is not the
only factor to be taken into account in assessing its productivity. Imagine
two firms which achieve the same output, but on the basis of different
inputs, i.e. at different production costs: the one with the lower
production costs would be considered the more productive. Production
cost is the second factor to be taken into account in assessing productivity.
It is a negative factor: other things being equal, the higher the production
cost, the lower the productivity.
Similar remarks apply to the assessment of relevance. The contextual

effects of an assumption in a given context are not the only factor to be
taken into account in assessing its degree of relevance. Contextual effects
are brought about by mental processes. Mental processes, like all
biological processes, involve a certain effort, a certain expenditure of
energy. The processing effort involved in achieving contextual effects is
the second factor to be taken into account in assessing degrees of
relevance. Processing effort is a negative factor: other things being equal,
the greater the processing effort, the lower the relevance.
In the last section we considered a definition of relevance expressed in
terms of necessary and sufficient conditions. That is, relevance was
defined as a classificatory concept. We also suggested that such a
definition, although not incorrect, missed the fact that relevance is also,
and more importantly, a comparative concept.!
Comparative concepts are best defined in terms of what might be called
'extent' conditions. Consider, for instance, the ordinary language concept
flexible. First, an object is not just flexible~ it is more or less so; in other
words flexibility is not just a classificatory but also a comparative concept.
Second, degrees of flexibility depend on (at least) two logically independent factors, reflected in the following definition:
(9) Flexibility
Extent condition 1: an object is flexible to the extent that it is easy to
bend.
Extent condition 2: an object is flexible to the extent that the shape it
can be bent into differs from its initial shape.
If an object can be bent at all, then conditions 1 and 2 are satisfied to some
extent, and conversely. These two extent conditions, therefore, logically
imply a necessary and sufficient condition: an object is flexible if and only
if it can be bent. Since this necessary and sufficient condition is implied by
definition (9), it need not be stated independently.
Definition (9) makes comparisons possible only in some cases: other
things being equal, if object Ais easier to bend than object B, then it is
125
more flexible; or, other things being equal, if object A can be bent further
than object B, then it is more flexible. But if A is easy to bend into a not
very different shape and impossible to bend any further, and B can be bent
only with difficulty but can then be bent much further, definition (9) does
not allow a comparative judgement to be made; and this seems to reflect
the limitations of ordinary usage. Incidentally, if we wanted to give an
adequate representation of the logical entry of the ordinary language
concept of flexibility, we would reformulate the extent conditions of
definition (9) as inference rules, which could be done in several different
ways. But our reason for discussing degrees of flexibility is not to shed
light on ordinary comparative concepts; it is to illustrate the form that a
theoretical comparative concept might take.
We are trying to develop a theoretical concept of relevance, for use in
the study of communication and cognition. We expect this theoretical
concept to help predict people's intuitions, but not necessarily their use of
the word 'relevance' or of similar ordinary language terms. We can
improve on definition (7) of relevance by adopting an extent-conditions
format of the type just illustrated:
(10) Relevance
Extent condition 1: an assumption is relevant in a context to the extent
that its contextual effects in this context are large.
Extent condition 2: an assumption is relevant in a context to the extent
that the effort required to process it in this context is small.
This definition implies the necessary and sufficient condition of definition
(7), which therefore need not be stated independently.
The assessment of relevance, like the assessment of productivity, is a
matter of balancing output against input: here contextual effects against
processing effort. Definition (10) of relevance, like definition (9) of
flexibility, makes clear comparisons possible only in some cases: other
things being equal, an assumption with greater contextual effects is more
relevant; and, other things being equal, an assumption requiring a smaller
processing effort is more relevant.
Let us now illustrate this comparative notion of relevance with a few
artificial examples; artificial in particular in the sense that the contexts we
are using are much smaller and more arbitrary: than contexts used in reallife comprehension. Readers should try to resist the natural tendency to
supply much richer and more appropriate contexts, a tendency which will
be discussed at length later on.
Consider a context consisting of assumptions (l1a-c):
(11) (a) People who are getting married should consult a doctor about
possible hereditary risks to their children.
126
Relevance
(b) Two people both of whom have thalassemia should be warned
against having children.
(c) Susan has thalassemia.
Consider the effects that assumptions (12) and (13), both by hypothesis
equally strong, would have in this context:
(12) Susan, who has thalassemia, is getting married to Bill.
(13) Bill, who has thalassemia, is getting married to Susan.
Both (12) and (13) have some contextual effects in context (11), and are
therefore relevant by definition (10). In particular, both (12) and (13) carry
the contextual implication (14):
(14) Susan and Bill should consult a doctor about possible hereditary risks
to their children.
This corresponds to a first intuition that both assumptions are relevant in

this context.
However, there is a further intuition that, in this context, (13) is more
relevant than (12). We can account for this on the basis of definition (10).
In this context, (13) has a contextual implication which (12) lacks:
(15) Susan and Bill should be warned against having children.
But what about processing effort? (12) and (13) have the same
conceptual structure, and thus make the same deductive rules available to
the deductive device. By hypothesis, they are also being processed in the
same context. May we assume, then, that they require the same processing
efort? Yes, but only after having clarified a point.
Having to write down and process the contextual implication (15)
involves some processing effort. This effort will have to be made in the
processing of (13), which carries implication (15), but not in the
processing of (12), which does not. However, this processing effort is
unavoidable if any contextual effect at all is to be achieved. If the benefits
of achieving a contextual effect were never enough to offset the cost of the
processing effort needed to implement it, then a positive degree of
relevance could never be achieved. Thinking would not be worth the
effort.
Except when they are in a state of utter exhaustion, humans find
thinking worth the effort. We can therefore draw the empirical conclusion
that the processing effort needed simply to write down, a contextual
implication or to raise or lower the strength of an assumption is not
enough to offset the contribution thereby made to relevance. Moreover,
since this processing effort is always in proportion to the effects it
implements, it can be altogether ignored in assessments of relevance.
Presumably the mind itself only worries about avoidable processing
Def!.rees of relevance
127
effort. We too will consider only processing effort which results in a

contextual effect, and will discount processing effort which results from
the fact that a contextual effect llas been obtained.
With this clarification, we can now say that (12) and (13) take exactly the
same processing effort when they are processed in the same context.
Moreover, since (13) has greater contextual effects than (12) in the context
(l1a--c), our definition predicts that it should be more relevant, and this
prediction is intuitively correct.
To illustrate now how relative relevance is affected by processing effort,
compare (13) and (16):
(13) Bill, who has thalassemia, is getting married to Susan.
(16) Bill, who has thalassemia, is getting married to Susan, and 1967 was a
great year for French wines.
When (13) and (16) are processed in the context (11a-c), they have exactly
the same contextual effects: the extra information conveyed by (16) is
utterly unrelated to the context and has no contextual effect whatsoever.
However, this extra information requires some extra processing effort:
(16) introduces more conceptual material, and hence more deductive rules
and matching procedures. By our definition of relevance, (16) should thus
be less relevant than (13), which achieves the same contextual effects with a
smaller processing effort. This prediction is again intuitively correct.
The examples discussed so far have involved only one kind of
contextual effect: contextual implication. Let us now take an example
where different kinds of contextual effect are simultaneously achieved.
Consider context (17a-g), where the strength of each assumption is
indicated on the right:
Peter is richer than Sam.
Sam is richer than Bill.
Bill is richer than Jim.
Jim is richer than Charles.
Sam is richer than Sue.
(f) Sue is richer than Jim.
(g) Sue is richer than Charles.
(17)(a)
(b)
(c)
(d)
(e)
[certain]
[certain]
[certain]
[certain]
[strong]
[very weak]
[strong]
Suppose that a hearer who has context (17a-g) in mind takes everything
the speaker says as certain. Suppose that the speaker is in a position to
assert either (18) or (19):
(18) Sue is richer than Jim.
(19) Sue is richer than Peter.
Intuitively, the assumption expressed by (19) is the more relevant, and is
128
Relevance
the one which, other things being equal, the speaker should choose to
express.
This is easily accounted for in terms of our definition of relevance.
Assumption (18) has only two contextual effects in the context (17a-g):
first, it raises the strength of (17f) from very weak to certain, since it is
identical in content to (17f) and is itself certain; second, it raises the
strength of (17g) from strong to certain, since (17g) is synthetically implied
by (17d) and (17), which are now both certain.
Assumption (19) has five contextual effects. It contextually implies (20)
and (21):
(20) Sue is richer than Sam.
(21) Sue is richer than Bill.
[certain]
[certain]
Assumption (20) contradicts assumption (17e), and since (20) is stronger

(certain versus strong), (17e) is erased from the memory of the deductive
device, a third contextual effect. Assumption (19) also raises the strength
of (17) and (17g) to certain, a fourth and fifth contextual effect. These last
two effects are identical to the only two effects achieved by (18).
Since (19) has greater contextual effects than (18), and since they both
require exactly the same processing effort (discounting, as we said, the
extra effort needed to implement the contextual effects themselves), then,
by our definition, (19) should be more relevant than (18), which is
intuitively correct.
Suppose now that the hearer accepts the assumptions expressed by (18)
and (19) merely as weak. Our definition predicts that in this case, (18)
should be more relevant than (19), reversing the previous order. The
hearer, who very weakly believed that Sue is richer than ]im, would have
his belief marginally independently strengthened from very weak to weak
by the assertion of (18). Thus (18) would achieve a modicum of relevance.
On the other hand, if the speaker expressed assumption (19), the hearer
would disbelieve it, since it contradicts his firm conviction that Sam is
richer than Sue. Assumption (19) would then be erased and would achieve
no relevance at all. This corresponds to the intuition that an unacceptably
exaggerated claim is irrelevant, while a modest and acceptable claim may
achieve some relevance, merely by confirming one's own assumptions.
Note, however, that the fact that an exaggerated claim is being made
may itself be relevant, which makes intuitions about such examples harder
to handle. For instance, the hearer might disbelieve (19), and nevertheless
reason that the speaker must have had some ground for believing that Sue
is rich; he might then see this as independently strengthening his own
assumptions (17f) and (17g). For this line of reasoning to be possible, the
context (17a-g) would have to be enriched, and in any case relevance
would be achieved on the basis not of assumption (19) but of assumption
(22):
129
(22) The speaker believes (19).

Still supposing that the hearer accepts the speaker's assertions only as
weak, consider (23):
(23) Sue is richer than Peter or she is richer than Jim.
The assumption expressed by (23) would have exactly the same contextual
effect as (18), i.e. it would strengthen (17f) from very weak to weak.
However, it would achieve this effect at a greater processing cost: a series
of deductive steps would be needed to reject the first disjunct of (23),
which is identical to (19), and accept as weak the second disjunct, which is
identical to (18). After this initial effort, the processing of (23) would be
the same as for (18). Our definition of relevance therefore predicts that (23)
should be less relevant than (18), a prediction which seems once more to be
intuitively correct. 2
Definition (10) of relevance does not make it possible to compare any
odd pair of assumptions in any arbitrary context. Imagine, for instance, a
very large and disparate context, consisting, say, of the contents of this
morning's Times, and two assumptions which both have substantial but
quite different contextual effects in that context. What procedure could be
used to compare the resulting contextual effects?
Or consider (24) and (25):
(24) Your garden will be a riot of colour in the spring if you plant these
beautiful bulbs.
(25) It has often been claimed that chlorinated water is less pleasant than
seawater to swim in.
What procedure could be used to compare the effort needed to process
these two assumptions, either in isolation from any context, or in any
selected context? More generally, could such procedures for assessing
contextual effects, processing effort and relevance be fully specified?
There is another way of putting essentially the same question: could
relevance be defined not just as a comparative but as a quantitative
concept? Our answer is: yes, it could. It is also quite conceivable that such
a quantitative notion of relevance would be of some interest to logicians.
However, it is not the kind of notion psychologists should be trying to
develop.
At an abstract level, the notion of relevance applies not just to human
beings, but to any information-processing device which is not simply
involved in achieving a fixed goal at a fixed cost. For instance, one might
want to characterise relevance for some abstract automaton. Suppose our
automaton is capable of achieving only one kind of contextual effect,
namely contextual implications. Then the contextual effects achieved by
adding an assumption to a context could be measured by counting
130
Relevance
contextual implications. Contextual effects involving changes in confirmation value could also be measured, as long as these values too were
quantitative, i.e. of the kind favpured by logicians.
Suppose further that all the operations of our automaton can be
analysed as combinations of equally simple elementary operations; in this
case the processing effort needed for a certain ta~k, e.g. achieving certain
contextual effects, could be measured by counting the elementary
operations involved. Or, if the automaton were implemented in the form
of a computer program, processing effort could be measured in terms of
the time taken to achieve particular effects. Then it would just be a matter
of deciding, in a principled or arbitrary way, h'ow contextual effects and
processing effort should be weighted against one another, and relevance
for this automaton could be quantitatively defined.
Things go differently when it comes to assessing contextual effects
achieved by human minds, and the processing effort needed to achieve
them. On the contextual-effects side, we have argued that nonquantitative confirmation values are involved. If so, then these effects
cannot be measured. On the processing-effort side, the prospects for
quantitative assessment are no better. For example, we do not know what
elementary operations complex thought processes reduce to. We do know
that the duration of a mental process is not an adequate indicator of its cost
for the organism: time spent in high mental concentration involves greater
effort than equal time spent in relaxed daydreaming.
The problems involved in measuring contextual effects and processing
effort are, of course, by no means specific to relevance theory or to
pragmatics. They affect psychology as a whole. However, for relevance
theory these problems take on a more specific form. Within relevance
theory, the problem is not so much to assess contextual effects and
processing effort from the outside, but to describe how the mind assesses
its own achievements and efforts from the inside, and decides as a result to
pursue its efforts or reallocate them in different directions.
Here is one line of possible speculation: contextual effects and mental
effort, just like bodily movements and muscular effort, must cause some
symptomatic physico-chemical changes. We might assume that the mind
assesses its own efforts and their effects by monitoring these changes.
Although we have nothing to say on the neuro-physics or neurochemistry involved, this is not an empty assumption. It contrasts with
another conceivable view/ on which contextual effects would beFassessed
by actually counting contextual implications, and processing effort by
actually counting inferential steps. There are many reasons for rejecting
this view: counting each step means adding one operation at each step,
which should considerably increase the effort involved in every mental
process. This in turn would be paradoxical, since presumably the point 6f
131
assessing effort is to be better able to reduce it. Moreover, if the assessment

of contextual effects and processing effort were the result of such a
computation, people should be able to make absolute judgements and
compare the contextual effects and processing effort involved in any pair
of mental performances, however unrelated; this expectation does not
seem to be borne out.
Contextual effects and processing effort are non-representational
dimensions of mental processes. They exist whether or not the individual
is consciously assessing them, whether or not they are conceptually
represented. When they are represented, we claim that they are represented in the form of comparative judgements. These judgements are
intuitive; they have their basis in the monitoring of physico-chemical
parameters.
People have not only retrospective intuitions about effects already
achieved and effort already incurred, but also prospective intuitions. That
is, they have intuitions about the effort some task would take and the
effects it might achieve (just as they have intuitions about the effort some
future bodily movement would take, and about its possible effects).
Prospective intuitions must be based not on the monitoring of physicochemical parameters, but on factors which systematically modify the
value of these parameters.
It is easy enough to identify a variety of factors that might make it
possible to predict which information will have the greatest contextual
effects. For instance, other things being equal, stronger assumptions have
greater contextual effects. Similarly, a variety of factors may make it
possible to predict how the processing effort needed to perform particular
pairs of tasks would compare. For instance, processing more information
in the same context, or the same information in a larger context, involves a
greater effort. People can take advantage of these comparative abilities in
trying to maximise the relevance of the information they process.
How are the two factors governing assessments of relevance balanced:
which effects are worth which effort? In a purely formal system, this
would be a matter for stipulation rather than discovery. In a computer
used for economic benefit, effort and effect could be assessed, say, in
dollars and cents. In the case of psychological processes, the problem
seems unamenable to any general solution, but then, on closer examination, it need not have a general solution at all.
It is extremely unlikely that the relative importance of effect and effort
stays constant across all circumstances and individuals. For instance,
changes in alertness may well alter one's willingness to incur a certain
processing effort: at some times the hope of achieving a given level of
contextual effect will suffice, and at others, not. Then, some people are
generally alert, and everything that is relevant at all will be more relevant
132
Relevance
for them than for duller people. Speakers who are not aware of their
hearers' disposition in the matter risk asking them for too much effort or
providing them with too few effects;
Mental effects and effort are non-representational properties of mental
processes. Relevance, which is a function of effect and effort, is a
non-representational property too. That is, relevance is a property which
need not be represented, let alone computed, in order to be achieved.
When it is represented, it is represented in terms of comparative
judgements and gross absolute judgements, (e.g. 'irrelevant', 'weakly
relevant', 'very relevant'), but not in terms of fipe absolute judgements, i.e.
quantitative ones.
Since we are interested in relevance as a psychological property, we have
no reason to aim for a quantitative definition of relevance. What we have
to do is add empirical substance to our comparative definition by
considering how relevance is sought and achieved in mental processes, and
particularly in processes of verbal comprehension. Our first task is to
move from a purely formal characterisation of a context to a more
empirical one, and to consider the implications of such a move.
3 Is the context given or chosen?
We have suggested that the context used to process new assumptions is,
essentially, a subset of the individual's old assumptions, with which the
new assumptions combine to yield a variety of contextual effects. We have
also proposed two criteria for comparing the relevance of different
assumptions in a given context. However, we still have to face the serious
problem of how the context is determined: how some particular subset of
the individual's assumptions is selected. For ease of exposition, we will
discuss this problem with reference to a particular case: that of a hearer
processing an assumption explicitly asserted by a speaker. In section 6, we
will generalise our account to deal with the assumptions made manifest by
any kind of stimulus.
In this section, we will look at various approaches which take for
granted that, at any given moment, there is only one context available to
the individual, and try to show that they fail precisely because of this
underlying hypothesis. In the next section we will suggest an alternative
approach.
In much of the literature, it is explicitly or implicitly assumed that the
context for the comprehension of a given utterance is not a matter of
choice; at any given point in a verbal exchange, the context is seen as
uniquely determined, as given. 4 Moreover, it is generally assumed that the
context is determined in advance of the comprehension process. The
Is the conte\xt given or chosen?
133
assumption explicitly expressed by an utterance is seen as combining with

a context present in the hearer's mind at the start of the act of utterance.
The simplest version of this view is the hypothesis that the context for the
comprehension of a given utterance is the set of assumptions explicitly
expressed by preceding utterances in the same dialogue or discourse. This
first hypothesis would seem to be borne out by the following exchange:
(26) (a) Peter: I'm tired.
(b) Mary: If you're tired, I'll make the meal.
It is easy to imagine a situation in which Mary's answer would be
intuitively relevant. In a discourse context consisting of the assumption
expressed by Peter, the assumption expressed by Mary would contextually imply (27), and the fact that it is relevant would seem to be thereby
explained:
(27) Mary will make the meal.

However, consider another version of the dialogue:
(b) Mary: I'll make the meal.
Intuitively, there is very little difference between Mary's answer in (26)
and her answer in (28): both are relevant in more or less the same way . Yet
if the context for comprehension were just the assumption explicitly
expressed by Peter, we would have to treat Mary's two answers quite
differently: (28b), unlike (26b), has no contextual effect whatsoever in
such a context, and should therefore not be relevant at all.
Let us then consider, as a second hypothesis, the claim that the context
for comprehension contains not only all the assumptions explicitly
expressed by preceding utterances in the discourse, but also all the
implicatures of these utterances. We can reasonably assume that in a
situation where Peter's remark was relevant, it would have implicated
something like (29):
(29) Peter wishes Mary would make the meal.
With (29) as part of the context, both (26b) and (28b) contextually imply
(30):
(30) Mary will do what Peter wishes.
They will thus both be relevant in this context, and relevant in similar
ways. Although (26b) has two contextual implications while (28b) has
only one, this is offset by the fact that (26b) has a more complex logical
form than (28b), and needs more processing. The fact that the two answers
134
Relevance
are intuitively similar in relevance could thus be straightforwardly

explained.
Consider, however, a third version of the dialogue:
(b) Mary: The dessert is ready. I'll make the main course.
Neither of the hypotheses considered so far can account for the fact that
Mary's answer (31 b) is relevant in roughly the same way as her answers in
the two earlier versions of the dialogue. (31 b) has no contextual effect in a
context consisting of either the assumptions ~plicitly expressed in
previous discourse, or the assumptions previously expressed and implicated. To account for the relevance of (31 b), the context used by the hearer
would have to include a premise such as (32):
(32) A meal consists of at least a main course and a dessert.
With (32) added to the context, contextual implication (33) can be derived
from (31b):
(33) Mary will make the meal.
Then from (33) and (29) (Peter wishes Mary would make the meal),
contextual implication (30) (Mary will do what Peter wishes) can be
derived, just as it was derivable, in a more restricted context, from (26b) or
(28b).
It is plain common sense to assume that a contextual premise such as
(32) would be used in processing Mary's answer (31b). However, this is
incompatible with the hypothesis that the context for comprehension is
the set of assumptiOP-s expressed or implicated by previous utterances.
Peter's remark that he is tired does not assert or imply that a meal consists
of at least a main course and a dessert. Assumption (32) has to be specially
retrieved from the encyclopaedic entry of the concept of a meal.
We might then consider, as a third hypothesis, the claim that the c~~.~.~~!
for comprehension consists not only of the assumptions expiesse~ or
implicated by preceding utterances, but also of the encyclopaedic entries
attached to any concepts used in these assumptions. For instance, if Peter's
initial remark implicates that he wishes Mary would make the meal, then
the encyclopaedic entry of the concept of a meal, and in particular
assumption (32) (A meal consists of a main course and a dessert), is
automatically added to the context in which Mary's answer will be
interpreted. With this third hypothesis, the relevance of (31 b) is accounted
for.
However, consider a fourth version of the dialogue:
(b) Mary: The dessert is ready. I'll make an osso-bucco.
Is the context given or chosen?
135
Intuitively there is little difference between the relevance of Mary's fourth

answer (34b) and her third answer (31 b). The obvious way to account for
this is to assume that the context in which (34b) is interpreted contains an
assumption such as (35):
(35) An osso-bucco is a main course.
With (35) in the context, the assumption explicitly expressed by Mary's
third answer (31 b) ('The dessert is ready. I'll make the main course') is
contextually implied by her fourth answer (34b), which explains the
similarity in relevance of (31 b) and (34b).
However, assumption (35) belongs to the encyclopaedic entry of
the concept of an osso-bucco. This concept did not occur in the
assumptions either expressed or implicated by Peter; it is introduced for
the first time by Mary's answer. This is not compatible with the
hypothesis that the context for comprehension consists of the assumptions expressed or implicated by preceding utterances, together with the
encyclopaedic entries attached to any concepts used in these assumptions.
One might be tempted, then, to formulate a fourth hypothesis to
accommodate the fourth version of the dialogue: the context for the
comprehension of an utterance consists of the assumptions expressed and
implicated by preceding utterances, plus the encyclopaedic entries
attached to any concept used in any of these assumptions, plus the
encyclopaedic entries attached to any concept used in the new utterance.
Note that on this hypothesis the context, though uniquely determined, is
not fixed in advance of the comprehension process. This fourth hypothesis implies, instead, that one of the preliminary stages of comprehension
consists in identifying the concepts used in the new utterance and adding
their encyclopaedic entries to the context. H0wever, there is still no
question of a choice of contexts.
With this fourth hypothesis, (35) is part of the context in which Mary's
answer (34b) is interpreted, and the relevance of her answer is thus
explained.
Consider, however, a fifth version of the dialogue:
(b) Mary: The dessert is ready. I'll make the speciality of the Capri
restaurant.
To establish the relevance of Mary's answer (36b), the hearer would have
first to access the encyclopaedic entry for the Capri restaurant and find out
that its speciality is osso-bucco, and then to access the entry for
osso-bucco and find out that an osso-bucco is a main course, i.e.
assumption (35). However, the concept of an osso-bucco occurs neither in
Mary's answer, nor in the assumptions expressed or implicated by Peter's
136
Relevance
initial remark. So, according to our fourth hypothesis, (35) is not part of
the context for the interpretation of (36b).
This might lead us, if we still had the stamina, to formulate a fifth
hypothesis: the context for the comprehension of an utterance consists of
the assumptions expressed and implicated by preceding utterances, plus
the encyclopaedic entries attached to any concepts used in these assumptions and in the utterance itself, plus the encyclopaedic entries attached to
any concepts used in the assumptions contained in the encyclopaedic
entries already added to the context. With our fourth hypothesis, one
layer of encyclopaedic entries was added to the context. With our fifth
hypothesis, two layers are added.
The defects of this line of speculation are becoming blatant. With the
last two hypotheses, we have already assumed that the context is
automatically filled with a huge amount of encyclopaedic information,
most - and sometimes all- of which fails to increase the contextual effects
of the new information being processed. Since each expansion of the
context means an increase in processing effort, this method of context
formation would lead to a general loss of relevance. Imagine the following
dialogue, for instance:
(37) (a) Peter: Where does John live?
(b) Mary: John lives next to the Capri restaurant.
If our fourth hypothesis were correct, the context in which Peter

interpreted Mary's answer should include the information that the
speciality of the Capri restaurant is osso-bucco. If our fifth hypothesis
were correct, it should also include the information that an osso-bucco is a
main course. This would be of no use - indeed it would be a distraction - in
understanding wh~re John lives.
Moreover, it is easy enough to find examples showing that two layers of
encylopaedic information may not be en<1Ugh. Suppose that in ~nswer to
Peter's remark that he is tired, Mary says,
(38) The dessert is ready. I'll make the speciality of that restaurant next to
where John lives.
"
First the encyclopaedic entry for John (and the information that he lives
next to the Capri restaurant) would be added to the context, which would
cause the entry for the Capri restaurant (and the information that its
speciality is osso-bucco) to be added. However, the hearer would still
need the information that an osso-bucco is a main course, and that
information is to be found in the entry for osso-bucco. To ensure that the
entry for osso-bucco is part of a uniquely determined context, three layers
of encyclopaedic information. would have to be automatically added.
Other examples would show that more and more layers of encyclopaedic
137
information might be needed. Soon, the context would be shown to

consist of the whole of encyclopaedic memory.
If the context included the whole of the hearer's encyclopaedia,
virtually any new information that a speaker could express would be
relevant, since virtually any new information would have some contextual
effects in such an enormous context. On the other hand, given the size of
such a context, an enormous processing effort - not to mention processing
time - would be needed to achieve these effects. Since relevance decreases
as more effort is required, this would mean that, while any new
information would easily achieve relevance, no information would ever
achieve more than minimal relevance. Furthermore, reminders could
never be relevant, since a reminder, on this approach, would merely be
repeating information already included in the context. This line is clearly
not worth pursuing.
Up to now, we have accepted the widely held view that the context in
which a given assumption is to be interpreted is uniquely determined. We
have seen the context as being formed either before the comprehension
process gets under way, or as a preliminary stage in this process. As we
have tried to show, assuming that the context is uniquely determined leads
to absurdities. However, there is nothing in the nature of a context, or of
comprehension, which excludes the possibility that context formation is
open to choices and revisions throughout the comprehension process. In
the next section, we explore this possibility further.
4 A choice of contexts
.)
In the last section, we talked rather generally of the context for

comprehension. Let us now be more specific and try to add some
psychological substance to the notion of a context introduced in chapter 2.
The set of assumptions in the memory of the deductive device at the start
of a deductive process can be partitioned into two proper subsets, each
acting as the context in which the other subset is processed. So far, this is a
purely formal move, It enables us to single out those synthetic implications whose, derivation actually involves both subsets of assumptions,
and describe them as contextual implications of one subset of assumptions
in the context of the other. It can then be used to clarify the more
psychologically significant distinction between information in the forefront of attention, which is usually new, and information which is taken
into account but remains in the background of attention, and which is
usually old: a distinction which typically applies to ordinary inferential
processes.
We assume that a crucial step in the processing of new information, and
138,
Relevance
in particular of verbally communicated information, is to combine it with

an adequately selected set of background assumptions - which then
constitutes the context - in the m:emory of the deductive device. For each
item of new information, many different sets of assumptions from diverse
sources (long-term memory, short-term memory, perception) might be
selected as context. However, this is not to say that any arbitrary subset of
the total set of assumptions available to the organism might become a
context. The organisation of the individual's encyclopaedic memory, and
the mental activity in which he is engaged, limit the class of potential
contexts from which an actual context can be chosen at any given time.
For instance, it is generally agreed that encycl~paedic information in
long-term memory is organised into chunks of some kind. Such chunks
have been discussed in the literature under such names as 'schema',
'frame', 'scenario' and 'prototype'. The encyclopaedic entries we have
mentioned are also chunks of a certain size, which may themselves be
grouped into larger chunks, and contain smaller chunks. It seems
reasonable to assume that the smallest units which can be transferred from
encyclopaedic memory to the memory of the deductive device are chunks
rather than individual assumptions. For instance, you might be unable to
remember and add to the context the information that the speciality of the
Capri restaurant is osso-bucco without also remembering and adding
other pieces of information about that restaurant: say, that the house's red
wine is a Valpolicella.
Moreover, not all chunks of encyclopaedic information are equally
accessible at any given time. We have no precise and well-grounded theory
of conceptual information retrieval, but various plausible assumptions
come to mind. It could well be, for instance, that the encyclopaedic entry
of a concept becomes accessible only when that concept appears in an
assumption that has already been accessed. For example, you might be
unable to recall that the speciality of the Capri restaurant is osso-bucco
unless you are already thinking of that restaurant (or of osso-bucco).
There will be times, then, when this information will be accessible in a
single step, times when it will be accessible in several steps, each'involving
an extension of the context, and times when the number of steps involved
will, in practice, make this information inaccessible.
,.
Consider someone about to process some new information. He still has
in mind some of the assumptions he has just been processing. People do
not come to the processing of new information with a 'blank mind'; they
have some kind of short-term memory store (or several such stores, or
devices functionally equivalent to short-term memory stores) whose
contents are never simply erased, at least not when the individual is awake.
However, it is not enough to point out that information may be carried
over from one conceptual process to the next; one would like to know
139
which information is kept in a short-term memory store, which is

transferred to encyclopaedic memory, which is simply !!rased. Here we
have neither formal arguments nor empirical evidence for any particular
set of hypotheses. Nevertheless, for the sake of concreteness and
explicitness, we will make a few suggestions, in particular about the
memory of the deductive device, which are compatible both with the little
that is understood and with the theory we are trying to develop.
At the start of each deductive process, the memory of the deductive
device contains an initial set of assumptions: that is, a set of premises. Then
all the non-trivial implications derivable from this set of premises are
derived, and all the strengthenings which can take place do. At the end of
the process, if no contradiction has resulted, the memory of the deductive
device contains all the original premises, possibly strengthened, and all the
newly derived conclusions. What happens now to these assumptions? We
will assume that all the newly derived synthetic implications, all the
premises which have undergone a synthetic rule, and all the premises
which have been strengthened, remain in the memory of the deductive
device; other assumptions still in the memory of the deductive device at
the end of the deductive process - that is, premises which have not affected
the deduction, or been affected by it - are erased from the memory of the
device. This is not to say, however, that they are not kept for a time in
some other short-term memory store.
There is good reason to think that the memory of the deductive device is
not the only short-term memory store available. Consider the fact that
someone may divide his attention between two tasks: say, watching
television and discussing family affairs at the same time. When this
happens, it seems that he may be switching bask and forth between two
quite different contexts. 5 This strongly suggests the existence of some
short-term conceptual memory other than that of the deductive device, in
/ which the context temporarily not in use is stored. We will assume that
assumptions erased from the memory of the deductive device are kept for
a time in such a general-purpose short-term memory store.
The assumptions left over in the memory of the deductive device from
the immediately preceding deductive process then constitute an immediately given context in which the next new item of information may be
deductively processed.
More particularly, a hearer who has just interpreted one utterance and is
about to interpret the next is characterised by the following distribution of
information. He has in the memory of his deductive device the set of
assumptions which make up his interpretation of the previous utterance
(including the premises actually used in achieving it). Other assumptions which were in the memory of his deductive device, but which have
played no role in the interpretation of the previous utterance, have now
140
Relevance
been transferred to his general short-term memory store. Also in thaJ

store, he has the interpretation of some earlier utterances (as well as other
items of information and thoughts he may have been attending to in the
immediate past). Possibly, part of the interpretation of the previous
utterance and of earlier utterances has been copied into encyclopaedic
memory. The assumptions left over in the memory of the deductive
device, i.e. the interpretation of the previous utterance, constitute an
immediately given context in which the next utterance may be processed.
We want to argue, however, that this immediately given context is
merely an initial context which can be extended in different directions.
Extensions of the context can be made by 'going back in time' and
adding to it assumptions used or derived in previous deductive processes.
The fact that such assumptions are easily accessible, which is introspectively and experimentally well established, is a further reason for thinking
that they are kept for a time in a general short-term memory store.
In the case of verbal understanding, the hearer may have to include in
the context not only the interpretation of the immediately preceding
utterance, but also the interpretation of utterances occurring earlier in the
exchange. Consider another version of the dialogue between Peter and
Mary:
(39) Mary: What I would like to eat tonight is an osso-bucco. I'm
ravenous. I had a great day in court. How was your day?
Peter: Not so good. Too many patients, and the air conditioning was
out of order. I'm tired.
Mary: I'm sorry to hear that. O.Ko I'll make it myself.
To understand Mary's concluding remark that she will make 'it' herself,
Peter needs information provided by her opening remark that she would
like to eat an osso-bucco. However, his interpretation of this opening
remark will in the meantime have been transferred from the memory of his
deductive device to his general short-term memory, if our above
hypotheses are correct. This interpretation must therefore be transferred
back to the memory of the deductive device, thus extending the
immediately given context (which consists of what is left in the memory of
Peter's deductive device after he has interpreted Mary's penultimate
remark that she is sorry to hear that he has had a bad day).
A second way of extending the context is to add to it the encyclopaedic
entries (or possibly smaller chunks of encyclopaedic information, taken
from these entries) of concepts already present either in the context or in
the assumption being processed. We have shown the need for such
extensions with examples (31 )-(38) above. We have also shown that the
assumption that such encyclopaedic extensions are automatically made
for every concept and in every case leads to absurdities; we used this as an
141
argument against the view that the context is uniquely determined. On the
other hand, once the determination of a context is seen as a matter of
choice and as part of the interpretation process itself, it seems reasonable
to assume that such extensions take place when they appear to be neededand only then.
A third way of extending the context is to add to it information about
the immediately observable environment. People constantly monitor the
physical environment while carrying out conceptual tasks which may be
partly or totally unrelated to that environment. Where is this subattentively monitored information stored? Again, we do not know, but
we can speculate: all this information is very briefly retained in specialised
short-term perceptual memory stores, from which some of it can be
transferred to the general short-term conceptual memory store and to the
memory of the deductive device. This happens, in particular, when the
interpretation of an utterance leads the hearer to pick up some environmental information and add it to the context. For example, suppose that
Mary, holding up a piece of veal, says to Peter,
(40) If you're tired, I'll cook this.
Peter will have to add to the context some description of the object Mary is
holding. The very form of Mary' s utterance provides an incentive to do so:
just as anaphoric pronouns, such as 'it' in (39), suggest going back in
discourse, deictic pronouns, such as 'this' in (40), suggest adding
environmental information to the context.
We have so far suggested that the choice of a context for inferential
processes in general, and for comprehension in particular, is partly
determined at any given time by the contents ef the memory of the
, deductive device, those of the general-purpose short-term memory store,
,;' and those of the encyclopaedia, and by the information that can be
immediately picked up from the physical environment. These factors
determine not a single context but a range of possible contexts. What
determines the selection of a particular context out of that range? Our
answer is that the selection of a particular context is determined by the
search for relevance.
In much of the pragmatic literature, events are assumed to take place in
the following order: first the context is determined, then the interpretation process takes place, then relevance is assessed. In other words,
relevance is seen as a variable to be assessed in function of a predetermined context. However, from a psychological point of view, this is
a highly implausible model of comprehension. Humans are not in the
business of simply assessing the relevance of new information. They try to
process information as productively as possible; that is, they try to obtain
from each new item of information as great a contextual effect as possible
142
Relevance
for as small as possible a processing effort. The assessment of relevance is

not the goal of the comprehension process, but only a means to an end, the
end being to maximise the relevance of any information being processed.
If this is true, it suggests a complete reversal of the order of events in
comprehension. It is not that first the context is determined, and then
relevance is assessed. On the contrary, people hope that the assumption
being processed is relevant (or else they would not bother to process it at
all), and they try to select a context which will justify that hope: a context
which will maximise relevance. In verbal comprehension in particular, it is
relevance which is treated as given, and conte~t which is treated as a
variable. 6 In this section, we have considered how the domain of this
variable is determined.
This change of perspective raises an obvious question: we have defined
relevance as a relation between a given assumption and a given context.
But when the context is not given, as we are now claiming is the case in
human understanding, how is the relevance of an assumption to be
defined? In answering this question, we will use our formal definition of
relevance in a context as the basis for a psychologically more appropriate
characterisation of relevance to an individual.
5 Relevance to an individual
At the end of each deductive process, the individual has at his disposal a
particular set of accessible contexts. This set is partly ordereg:each
conte:l(t(aE~rt fr0E!_th~j_nitiaISQ.ntextj-contains one or more smaller
contexts:~anleac1i context
from the maximafcohtexts Yiscontained
in orie-or-moreIarg/!~contexts. The set of accessible contexts is thus partly
ordered by the inclusion-relation. This formal relation has a psychological
counterpart: order of inclusion corresponds to order of accessibility. The
initial, minimal context is immediately given; contexts which include only
the initial context as a sub-part can be accessed in (one step and are
therefore the most accessible contexts; contexts which include the initial
context and a one-step extension as sub-parts can be accessed in two steps
and are therefore the next most accessible contexts, and so on. Notice a
point of crucial importance for relevance theory: just as processing an item
of information in a context involves some effort, so accessing a context
involves some effort. The less accessible a context, the greater the effort
involved in accessing it, and conversely.
Consider a new assumption A. This may be relevant in some, all or none
of the contexts accessible to an individual at a given time, depending on
whether some, all or none of these contexts already contain or imply a
(apart
143
token of A, and on the relativ,e strength of old and new tokens. Six
situations can be distinguished (the list is not exhaustive, but is representative enough for our present purposes):
(41)(a) A is already contained in (or implied by) the initial context, at
maximal strength. Then the new token of A is irrelevant in this
context, and in all the other accessible contexts too, since all these
contexts include the initial context. 1'0. this situation there is no
point in searching for relevance beyond the initial context, since
the search will be unproductive.
(b) A is contained in (or implied by) none of the accessible contexts;
however A has no contextual effect in any of them either. Then
again, A is irrelevant in all the accessible contexts, and there is no
point in extending the initial context in the search for relevance.
(c) A is contained in (or implied by) the initial context and all
accessible contexts, at less than maximal strength. Then an
independent strengthening of A by the new token will ensure its
relevance in all the accessible contexts. In this situation an
extension of the context will be justified as long as A has more
contextual effects in the extended context than in the initial
context, and the gain in contextual effects is not outweighed by
the greater effort needed to process A in the extended context.
(d) A is contained in (or implied by) none of the accessible contexts,
and has some contextual implications in the initial context. Then
A is relevant in all the accessible contexts in which it retains these
contextual implications. Here again an extension of the context
will be justified as long as it yields greater contextual effects, and
the increase in contextual effects is not outweighed by the increase
in processing effort required.
(e) A is contained in (or implied by) none of the contexts; it has no
contextual effect in the initial context but has some contextual
effect in some extensions of the initial context. Then A is relevant
in some of the accessible contexts. In this situation, no relevance
will be achieved unless the context is extended. Extensions should
follow the pattern laid down in (c) and (d).
(f) A is not contained in (nor implied by) the initial context, but is
contained (at maximal strength) in some of the larger accessible
contexts; A has contextualeffects in some of the contexts in which
it is not contained (which mayor may not include the initial
context). Then A is relevant in some of the accessible contexts, and
its relevance will be that of a reminder. A reminder is relevant only
in contexts which do not contain the information in question: its
144
Relevance
function is to make this information accessible at a smaller
pro~essing cost than would De needed to obtain it by successive
extensions of the contexC
Intuitively, as we will shortly illustrate, an assumption is relevant to an

individual in situations (41c), (41d), (41e) and (41). In situations (41c) and
(41 d) its relevance is immediately manifest; the difference between the two
situations is that relevance is achieved in (41c) by contextual strengthening, and in (41d) by contextual implication. In situation (41e), it takes an
extension of the context, and hence some effort, to bring out its relevance,
but ordinarily this effort is not consciously felt, and again the relevance is
immediately or almost immediately manifest. In situation (41), the
relevance is that of a reminder and is again immediately manifest.
We can now provide a classificatory definition of relevance to an
individual :a
(42) Relevance to an individual (classificatory)
An assumption is relevant to an individual at a given 'time if and only if
it is relevant in one or more of the contexts accessible to that
individual at that time.
However, for reasons discussed in section 2, we are less interested in
classificatory than comparative definitions of relevance. Just as we did for
relevance in a context, we will now characterise the comparative notion of
relevance to an individual in terms of effect and effort. On the effort side,
what has to be taken into account is not only the effort needed to process
an assumption in a given context, but also the effort needed to access that
context. For each of the contexts accessible to an individual, the effect and
effort involved, and therefore the relevance achieved, will differ. Indeed,
the same context can be accessible in different ways, involving different
amounts of effort and therefore different relevance values. We might
therefore try to characterise the relevance of an assumption to an
individual in terms of a set of relevance values, one for each possible way
of processing that assumption: i.e. one for each possible context and
method of accessing that context.
However, the result of this cumbersome procedure would be of little
psychological interest. We assume that the individual automatically aims
at maximal relevance, and that it is estimates of this maximal relevance
which affect his cognitive behaviour. Achieving maximal relevance
involves selecting the best possible context in which to process an
assumption: that is, the context enabling the best possible balance of effort
against effect to be achieved. When such a balance is achieved, we will say
that the assumption has been optimally processed. When we talk of the
relevance of an assumption to an individual, we will mean the relevance
achieved when it is optimally processed. We now define:
<
145
(43) Relevance to an individual (comparative)
Extent condition 1: an assumption is relevant to an individual to the
extent that the contextual effects achieved when it is optimally
processed are large.
Extent condition 2: an assumption is relevant to an individual to the
extent that the effort required to process it optimally is small.
As with definition (10) (the comparative definition of relevance in a
context), this definition of relevance to an individual does not make
comparisons possible in all cases. Take two unrelated assumptions, each
relevant to a different individual at a different time: is assumption A 1 more
relevant to Bill at time tl than assumption A2 is to Joan at time t2? Our
definition does not normally make it possible to answer such a question,
nor, from a psychological point of view, is there any reason why it should.
The only comparisons of relevance that play a psychological role are those
which are subservient to the goal of maximising relevance: relevance to
oneself, or, from the point of view of a communicator, relevance to an
audience.
Let us illustrate this definition of relevance to an individual with an
example somewhat less fragmentary than previous ones (although it still
does not come near the complexities of real-life information processing).
Suppose that the following exchange has so far taken place:
(44) Mary: What I would like to eat tonight is an osso-bucco.

Peter: I had a long day. I'm tired.
After interpreting Peter's last remark, let us 'assume that Mary has in the
memory of her deductive device an initial context composed of the three
assumptions (45a-c), which are strong but 30t certain:
(45) Initial context

(a) Peter is tired.
(b) If Peter is tired, he wishes Mary would make the dinner.
(c) Peter wishes Mary would make the dinner.
(45a) is the assumption expressed by the last utterance to be processed;
(45b) is a premise which combined with (45a) to yield the contextual
implication (45c). Other assumptions which Mary may have had in the
memory of her deductive device when she began interpreting Peter's last
remark have failed to yield contextual effects and have therefore been
erased at the end of the interpretation process. In real life, the set of
assumptions left in the memory of Mary's deductive device after she has
interpreted Peter's last remark would presumably be much larger.
Mary can extend this initial context by adding to it various chunks of
information, in particular:
146
Chunk 1.
Chunk 2.
Chunk 3.
Chunk 4.
Chunk 5.
Relevance
Encyclopaedic information about Peter, including the
assumption, Peter isa surgeon.
Encyclopaedic information about Mary.
Encyclopaedic information about making dinner, including a
scenario of looking in the refrigerator to see what is available,
and the assumption, A dinner consists of at least a main course
and a dessert.
.
Information about the currently monitored physical environment.
Assumptions processed at earlier. stages in the exchange,
including: Mary would like to eat an osso-bucco.
Chunks 1-5 are accessible in one step from the initial context. Each of
these potential extensions makes further extensions. accessible in turn. For
instance, the information that Peter is a surgeon makes chunk 6 accessible:
Chunk 6.
Encyclopaedic information about surgery.
The scenario of looking in the refrigerator in chunk 3 makes chunk 7

~ccessible:
Chunk 7.
What Mary remembers of what there is in the refrigerator,

including the assumption, There is a chocolate mousse in the
refrigerator.
The concept of an osso-bucco appearing in chunk 5 makes chunk 8

accessible:
Chunk 8.
Encyclopaedic information about osso-bucco, including the

two assumptions, An osso-bucco is a main course and An
osso-bucco is a veal dish.
Chunks 6, 7 and 8 are accessible only as a result of a two-step extension of

the initial context (45). They make other chunks of information accessible
in turn. For instance, chunk 9 is accessible as long as the concept of a
coronary bypass appears in chunk 6 (information about surgery):
Chunk 9.
Encyclopaedic information about coronary bypass, including

the assumption, Performing a coronary bypass is exhausting.
Similarly, chunk 10 is made accessible by the presence of the concept of a

chocolate mousse in chunk 7:
Chunk 10. Encyclopaedic information about chocolate mousse, including the assumption, A chocolate mousse is a dessert.
Of course, further levels of extension and many more extensions at each
level are possible, but we will stop here and consider what effect various
147
continuations of the dialogue between Peter and Mary might have on
context selection.
Case A. Suppose first that Peter stops after saying, 'I'm tired'. Mary might
then have a thought which is relevant to her in the context (45). She may
for instance decide to make the dinner herself, which contextually implies
that she will do what Peter wishes. This contextual implication makes her
decision relevant not only to her but also to Peter, so she might decide to
inform him of it aqd say,
(46) Mary: If you're tired, I'll make the dinner.

Mary may also extend the context to include chunk 3 (information
about making dinner) and chunk 5 (in particular the information that she
would like to eat osso-bucco). She can then derive many more contextual
effects from her decision, in the form of more specific decisions about
what to cook, for instance an osso-bucco, and about the various practical
steps to take, like opening the refrigerator, etc.
This case has implications for the role of relevance in thought processes
in general, not just in the interpretation of utterances. Each thought
process leaves the mind in a state characterised by an initially given context
and possible extensions. If we are right in assuming that the train of human
thoughts is steered by the search for maximal relevance, then the mind
should try to pick out, from whatever sources. it has available, including its
own internal resources, the information which has the greatest relevance
in the initial context: that is, which has the greatest contextual effects and
requires the smallest processing effort. Such information is to be sought in
accessible extensions of the context, whether they involve encyclopaedic
memory, the short-term memory store, or the environment. Thus,
relevance theory yields hypotheses about the way thoughts follow one
another, and about the points at which the individual might turn to the
environment, rather than to his own internal resources, for relevant
information.
Case B. Suppose that the dialogue (repeated for convenience with the new
development italicised) continues as in (47):
(47) Mary: What 1 would like to eat tonight is an osso-bucco.

Peter: 1 had a long day. I'm tired. I wish you would make the dinner.
Peter's last remark ('1 wish you would make the dinner') achieves
relevance in the initial context (45a-c) by strengthening the contextual
implication (45c) that Mary had derived from his preceding remark ('I'm
tired'). From then on, Mary's train of thought should proceed as it would
have if Peter had stopped after saying 'I'm tired': that is, along the lines
148
Relevance
considered in Case A. The relevance achieved should be greater, though,

since all conclusions based on premise (45c) (Peter wishes Mary would
make the dinner) will be strengthened. This is an illustration, then, of
situation (41c): some relevance is achieved in all accessible contexts by the
strengthening of an existing assumption.
Case C. Suppose that the dialogue continues as in (48):

Peter: I had a long day. I'm tired. If we're both tired, I'd like us to go
to the Capri restaurant instead of having to make dinner.
Peter's last remark has a contextual implication in the context (45a-c); it
contextually implies (49):
(49) If Mary is tired, Peter would like them to go to the Capri restaurant.
Peter's last remark achieves relevance in all accessible contexts because of
this contextual implication. This is therefore an illustration of situation
(41d). At the same time, his remark achieves relevance in all accessible
contexts in another way: by contradicting, and thereby eliminating,
assumptions (45b) (If Peter is tired, he wishes Mary would make the
dinner) and (45c) (Peter wishes Mary would make the dinner).
Peter's remark also makes accessible an extra chunk of information:
Chunk 11. Encyclopaedic information about the Capri restaurant, {ncluding the assumption, The speciality of the Capri restaurant
is osso-bucco.
Now clearly, some possible extensions of the context would diminish
overall relevance: for instance there would be no gain in contextual effects
from the addition of chunk 6 (information about surgery), and the extra
processing costs would lead to a loss of relevance. Other extensions,
however, would increase relevance. Suppose, for instance, that chunk 2
contains the assumption (50):
(50) Mary is tired.
With (50) added to the context, Peter's last remark contextually implies
(51):
(51) Peter would like them to go to the Capri restaurant.
Adding chunk 5, and in particular the information that Mary would like to
eat an osso-bucco, and chunk 11 (information about the Capri restaurant)
yields another contextual implication:
(52) Peter would like them to go to a restaurant the speciality of which is
what Mary would like to eat.
Relevance to.an individual
149
This leads in turn to many more contextual implications and strengthenings in a context containing information about Mary and .Peter (and about
osso-bucco ).
Case D. Suppose that the dialogue continues as in (53):
Peter: I had a long day. I'm tired. I've just done a coronary bypass.
Peter's last remark is not relevant in the initial context (45a-c). However,
it is relevant in a context extended to include chunk 9 (information about
coronary bypass, including the assumption that performing a coronary
bypass is exhausting). This extension, which was accessible in three steps
from the initial context, has now become accessible in one step thanks to
the presence of the concept of a coronary bypass in Peter's utterance. In a
context so extended, the assumption that Peter has just done a coronary
bypass contextually strengthens (45a) (Peter is tired), and achieves
relevance thereby. This, then, is an illustration of situation (41e).
Case E. Suppose that the dialogue continues as in (54):
Peter: I had a long day. I'm tired. I wish you would make the dinner
tonight, and, by the way, there's a dessert, a chocolate mousse, in the
refrigerator.
)
The first part of Peter's last remark ('I wish you would make the dinner
tonight') is relevant as described in the discussion of Case B.1t should lead
to the addition of chunk 3 (and in particular the assumption that a dinner
consists of at least a main course and a dessert) to the context, as described
in the discussion of Case A. This one-step extension makes chunk 7
(which contains the information that there is a chocolate mousse in the
refrigerator) accessible in one further step, which, in a third step, makes
accessible chunk 10 (which contains the information that a chocolate
mousse is a dessert).
From the assumptions that would be available in the memory of Mary's
deductive device if she carried out this three-step extension of the initial
context (45a-c), she would be able to deduce that all she has to do to make
the dinner is make a main course. The second part of Peter's last remark
('There's a dessert, a chocolate mousse, in the refrigerator') makes the
same conclusion available without her having to extend the context
beyond adding chunk 3. It also makes chunk 10 (information about
chocolate mousse) accessible in one step, without her having to go
through chunk 7 (the contents of the refrigerator).
Case E illustrates situation (41), and shows how a reminder may be
relevant: the effort needed to retrieve some relevant information from
150
Relevance
memory may be greater than the effort needed to get the same information
from the interpretation of an utterance. In these-circumstances, a reminder
is relevant: the contextual effects it produces could have been produced
otherwise, but more slowly and at a greater processing cost. On the other
hand, when a reminder comes after the context has been extended to
include the very information that the speaker is trying to bring to the
hearer's attention, then the extra effort needed to process an utterance
which carries no new contextual effects is wasted, and redundancy rather
than relevance is the result.
The five cases discussed above show how the ratker abstract notion of
relevance in a context can help with the construction of a psychologically
more significant notion of relevance to an individual. They also show the
crucial importance of the organisation of encyclopaedic memory in the
pursuit of relevance. In fact, the relation between memory and relevance is
so close that relevance theory might well shed new light on the
organisation of memory itself. For instance, the way in which information
is chunked may in principle help or hinder the search for relevance;
plausibly, forms of chunking which are a help rather than a hindrance tend
to predominate. Conversely, the pursuit of relevance may lead to the
faster building and enrichment of chunks of a certain form.
In this section we have characterised and illustrated a notion of
relevance to an individual. We have done this in an attempt to come closer
to a psychologically adequate notion of relevance, for use in describing
and explaining verbal comprehension and other cognitive processes. So
far, we have treated relevance as a property of assumptions. In particular,
we have equated the relevance of an utterance with the relevance of the
assumption it explicitly expresses. Yet hearers do not simply pick up the
assumption expressed by an utterance. More generally, individuals do not
simply pick up assumptions from their environment. In either case, a
complex cognitive process requiring mental effort is involved.
Conversely, a communicator cannot directly present an audience with
an assumption. All a speaker or any other type of communicator can do is
present a stimulus, hoping that its perception by members of the audience
will lead to a modification of their. cognitive environment and trigger
some cognitive processes. To the audience, a stimulus is initially just one
phenomenon among others: that is, just one perceptible feature of the
physical environment. It becomes identifiable as a stimulus only when it is
recognised as a phenomenon designed to achieve cognitive effects.
Which phenomena does the individual pay attention to? How does he
go about processing the information they make manifest? We want to
claim that he tends to pay attention to relevant phenomena, and to process
them so as to maximise relevance. However, to do this we have to
characterise relevance not just as a property of assumptions in the mind,
151
but also as a property of phenomena (stimuli, e.g. utterances) in the

environment which lead to the construction of assumptions .. This we will
do in the next section.
6 The relevance of phenomena and stimuli

An individual's cognitive environment is the set of all facts which are
manifest to him. A phenomenon affects the cognitive environment by
making certain facts manifest or more manifest. As a result, the individual
can mentally represent these facts as strong or stronger assumptions, and
perhaps use them to derive further assumptions which do not correspond
to actual facts, but which are nonetheless manifest to him too (see chapter
1, section 8, above).
A phenomenon may make manifest a very large number of assumptions. However, this is not to say that the individual will actually construct
any, let alone all, of these assumptions. The house has its usual smells; the
individual pays no attention to them and makes no assumptions about
them whatsoever. Now suppose there is a distinct smell of gas. The
individual is likely to make assumptions (55) and (56):
(55) There is a smell of gas.
(56) There is a gas leak somewhere in the house.
He is less likely to make assumption (57), even though it has become

manifest too:
(57) The gas company is not on strike.
Why does he make some assumptions and not others? First, there are
certain assumptions he cannot avoid making in a given cognitive
environment. Take auditory perception. The faculty of auditory perception handles a great number and variety of noises, few of which reach the
level of attention: that is, lead to the construction and manipulation of
conceptual representations by the central thought processes. The mechanisms of auditory perception act as a filter, processing and filtering out most
acoustic information at a sub-attentive level. These sub-attentively
processed phenomena may come to the individual's attention, but only
when central thought processes turn to the perceptual mechanisms for
information about them.
However, some acoustic phenomena automatically pre-empt attention,
automatically give rise to assumptions and inferences at a conceptual level.
The perceptual mechanisms are organised so as to let certain types of
phenomena impinge on central thought processes. Some of these favoured
types of phenomena are probably innately determined: for instance~ the
152
Relevance
automatic attention paid to all sudden loud noises has contributed to the
survival of the species and is presumably an outcome of natural selection.
Other types of phenomena pre-empt attentio Il as a result of some form
of learning. The crying of a particular baby, even if barely audible,
pre-empts the attention of the parents. A smell of gas pre-empts the
attention of gas-users. Once the individual has smelled the gas, he cannot
help but make assumption (55), the assumption that there is such a smell.
The automatic filtering out of some phenomena and the automatic
pre-empting of attention by others can be seen as a heuristic device aimed
at maximising cognitive efficiency: in general, it is the phenomena which
are least likely to be relevant which get filtered out, and those most likely
to be relevant which pre-empt attention. In other words, the perceptual
mechanisms - and perceptual salience itself - are relevance-oriented.
Assumption (56), the assumption that there is a gas leak, is a contextual
implication of assumption (55) in a context containing ordinary encyclopaedic information about household uses of gas. We want to suggest that
assumption (56) is made in an attempt to maximise the relevance of
assumption (55); indeed it is particularly useful in this respect, since it
gives easy access to many other contextual effects. Precisely because the
processing of (55) is governed by the search for relevance, assumption (57)
is unlikely to be made: the processing effort needed to derive (57) is greater
than the effort needed to derive (56), and moreover (57) does not lead to
rich contextual effects achievable at a low processing cost.
A phenomenon can be more or less efficiently processed depending on
which, if any, of the assumptions it makes manifest are actually
constructed. For some phenomena, the best course is to filter them out at a
perceptual level. For others, it is to represent them conceptually and
process them in a rich encyclopaedic context. The notion of relevance can
thus be extended to phenomena in a straightforward way:
(58) Relevance of a phenomenon (classificatory)
A phenomenon is relevant to an individual if and only if one or more
of the assumptions it makes manifest is relevant to him.
A comparative definition is similarly straightforward. As always, we
will characterise the comparative notion of relevance in terms of effect and
effort. Here, what has to be taken into account on the effort side is not
only the effort needed to access a context and process an assumption in
that context, but also the effort needed to construct that assumption. The
construction and processing of different assumptions will involve different effects and amounts of effort, and hence different degrees of relevance.
For reasons discussed in the last section, we will characterise the relevance of
a phenomenon to an individual as the relevance achieved when it is
optimally processed. 8
153
We now define:
(59) Relevance of a phenomenon (comparative)

Extent condition 1: a phenomenon is relevant to an individual to the
extent that the contextual effects achieved when it is optimally
Extent condition 2: a phenomenon is relevant to an individual to the
extent that the effort required to process it optimally is small.
A stimulus is a phenomenon designed to achieve cognitive effects.
Relevance for a stimulus is thus the same as relevance for any other
phenomenon, and definitions (58) and (59) apply directly. We have argued
that the processing of phenomena in general, and hence of stimuli in
particular, is geared to the maximisation of relevance. Someone who wants
to achieve a specific cognitive effect must therefore try to produce a
stimulus which, when optimally processed, will achieve just the intended
effect. This effect may be achieved at either the attentive or the
sub-attentive level. When a child wants her parents to feel sorry for her,
the best course might be to cry in a manifestly sincere way: the parents'
attention will be pre-empted, and the most relevant assumption will be
that the child is distressed.On the other hand, suppose Peter wants Mary
to feel aroused by the manly smell of his after-shave, but is afraid she will
be put off if she guesses his intention; his best (:ourse would be to use it
sparingly, since a strong smell might attract her attention and make his
intention all too manifest.
Here we are interested in stimuli used to achieve rather subtler cognitive
effects: stimuli used to make an informative intention mutually manifest.
Ostensive stimuli, as we will call them, must satisfy two conditions: first,
they must attract the audience's attention; and second, they must focus it
on the communicator's intentions.
Ostensive-inferential communication cannot achieve its effect subattentively; this necessarily involves the construction of conceptual representations and the mobilisation of central thought processes. This is why
most stimuli used in ostensive communication are attention-pre-empting:
they typically involve sudden loud noises such as shouts or doorbell
chimes, striking visual stimuli such as hand waves, flashing lights or bright
posters, or vigorous tactile stimulation such as prodding or grasping. Most
important of all, spoken utterances in one's own native language
automatically impinge on the attention: if they are distinctly audible, it is
almost impossible to filter them out as background noise. It is only when
the audience is likely to pay attention to the ostensive stimulus of its own
accord, as you are now doing, that the stimulus, little dark marks on white
paper for instance, can be a poor attractor of attention.
The second condition that an ostensive stimulus must meet is to focus
154
Relevance
the attention of the audience on the communicator's intentions. That is,
the assumption that the stimulus is ostensive must be both manifest
enough and relevant enough to lead tooptimal~processing. This condition
is generally met by stimuli which both pre-empt the attention and are
irrelevant unless treated as ostensive stimuli. This is clearly true of coded
signals used in ostensive communication, linguistic utterances in particular, which, unless treated as ostensive stimuli, are mere irrelevant noises or
marks on paper. It is also true of non-coded ostensive stimuli.
A non-coded ostensive stimulus may be an ordinary bodily movement,
with little intrinsic relevance, made with artificial- and attention-arresting
- rigidity: as when Peter leans back ostensivety to let Mary see William
coming (see chapter 1, section 9). It may be a piece of mimicry: for
instance, Mary might mimic the act of driving to communicate to Peter
that she wants to leave the party. Most of what such performances make
manifest is of little or no relevance. Someone has made some quite
ordinary bodily movement: so what? The only relevant assumptions
made manifest by such behaviour are assumptions about the individual's
informative intention.
The best ostensive stimuli are entirely irrelevant unless they are treated
as ostensive. Consider a case where an intrinsically highly relevant
stimulus is used - or misused - ostensively: say, somebody who is believed
to have her arms paralysed mimics the act of driving. Here, the fact that
she can move her arms would be so much more relevant than anything she
might have wanted to communicate that her informative intention might
well go unnoticed. Or to take a political example, acts of terrorism
designed to publicise a cause have so many important implications
irrespective of the terrorists' informative intention that they are much
better at attracting public attention than at conveying the intended
message.
However, it is not enough for the ostensive stimulus to attract attention
and focus it on the communicator's intentions. It must also reveal the
communicator's intentions. How can it do this? We will argue that what is
crucial here is that an ostensive stimulus comes with the communicator's
guarantee of relevance. b In general, there is no guarantee that a phenomenon will turn out to be relevant. Some phenomena are not relevant at
all, and are therefore not worth processing at a conceptual level; others
may be highly relevant, and may set off a whole train of thought. There
can be no a priori expectation of relevance for phenomena in general.
In the special case of ostensive stimuli, the situation is quite different.
By producing an utterance, the speaker requests her hearer's attention. By
requesting his attention, she suggests that her utterance is relevant enough
to be worth his attention. This applies not just to speech but to all forms of
ostensive communication. Ostensive stimuli arouse definite expectations
155
of relevance, of relevance achievable once the communicator's informative
intention is recognised. In the next section, we will develop this idea and
formalise it as a principle of relevance. Then, in the last section of this
chapter, we will show how the principle of relevance explains ostensiveinferential comunication.
7 The principle of relevance

We ended chapter 1 with the following definition of ostensive-inferential
communication:
(60) The communicator produces a stimulus which makes it mutually
manifest to communicator and audience that the communicator
intends, by means of this stimulus, to make manifest or more
manifest to the audience a set of assumptions I.
As we said, this definition does not explain how ostension works: how the
ostensive stimulus makes manifest the communicator's informative
intention. We suggested that an answer to this question was to be sought
in a principle of relevance, but that such a principle would not be truly
explanatory until the notion of relevance had itself been explicitly
characterised. Having done this, we can now return to the principle of
relevance.
To achieve its effect, an act of ostensive communication must attract
the audience's attention. In that sense,9 an act of ostension is a request
for attention. Someone who asks you to behave in a certain way, either
physically or cognitively, suggests that he has good reason to think it
might be in your own interests, as well as his, to comply with his request.
This suggestion may be ill founded or made in bad faith, but it cannot be
wholly cancelled. If a request has been made at all, the requester must have
assumed that the requestee would have some motive for complying with
it. Even a blackmailer has to make it look preferable for his victim to
co-operate rather than to refuse; similarly, when a drowning man calls for
help, his only chance is that some passer-by will find it morally preferable,
however physically inconvenient, to help him.
Less dramatically, the host who asks his guests to eat automatically
suggests that what he is offering them is edible, and indeed worth eating.
Just as feeding someone normally requires the participation of the
recipient in the form of appropriate bodily behaviour, ostensive communlcation requires the participation of the recipient in the form of
appropriate cognitive behaviour, and in particular of attention. If Mary
requests Peter's attention by pointing to something in the landscape, or
holding something up for him to see, or talking to him, he is entitled to
156
Relevance
assume that the stimulus being drawn to his attention is relevant to him,
or at least that she has reason to think it is; if she gives him something to
think about, she must believe that he will find it good food for thought.
There is thus a substantial difference between the frame of mind in
which the individual may approach an ostensive stimulus directed at him
and the frame of mind in which he approaches other phenomena. When
attending to other phenomena, he may have hopes of relevance: if such
hopes were totally unwarranted, there would be no point in attending to
them at all. However, whether these hopes turn out to be justified depends
on a variety of factors, most of which are beyond \he individual's control,
and which he may not even be aware of. What makes these hopes
reasonable is that humans have a number of heuristics, some of them
innate, others developed through experience, aimed at picking out
relevant phenomena. Even so, hopes of relevance sometimes turn out to be
unjustified, and when they are justified, they are justified to a greater or
lesser extent: there can be no general expectation of a steady and
satisfactory level of relevance.
With an ostensive stimulus, however, the addressee can have not only
hopes, but also fairly precise expectations of relevance. It is manifest that
an act of ostensive communication cannot achieve its effect unless the
audience pays attention to the ostensive stimulus. It is manifest that people
will pay attention to a phenomenon only if it seems relevant to them. It
is manifest, then, that a communicator who produces an ostensive stimulus
must intend it to seem relevant to her audience: that is, must intend to
make it manifest to the audience that the stimulus is relevant. Adding a
layer of mutuality to this account, let us suppose that it is not merely
manifest but mutually manifest to communicator and audience that an
ostensive stimulus is being produced. Then it is not merely manifest but
mutually manifest that the communicator must intend the stimulus to
seem relevant to the audience: that is, must intend it to be manifest to the
audience that the stimulus is relevant. By our definition of ostensiveinferential communication, this amounts to saying that an ostensive communicator necessarily communicates that the stimulus she uses is relevant
to the audience. In other words, an act of ostensive communication
automatically communicates a presumption of relevance.
What is the exact content of the presumption of relevance communicated by an act of ostensive communication? As we have said, what is
communicated is that to the best of the communicator's knowledge, the
ostensive stimulus is relevant enough to be worth the audience's attention.
No weaker guarantee would do. But the presumption of relevance is more
specific than this. The relevance of a stimulus is determined by two
factors: the effort needed to process it optimally, and the cognitive effects
157
this optimal processing achieves. We want to argue that the presumption

of relevance is different on the effect and effort sides. On the effect side,
the presumption is that the level of effects achievable is never less than is
needed to make the stimulus worth processing; on the effort side, it is that
the level of effort required is never more than is needed to achieve these
effects.
The communicator intends to communicate a set of assumptions I.
Of course, it is in the addressee's interest that I should be the most
relevant information available to the communicator. However, here the
interests of communicator and addressee need not coincide. The communicator may want to keep to herself the ~ost relevant information at
her disposal; she may have reasons of her own for communicating
information that is less relevant. A communicator wants to communicate
not just any arbitrary set of assumptions, but some particular set of
assumptions I, which she may have her own reasons for wanting to
convey. However, given that she needs the addressee's attention, she
cannot but communicate that I is relevant enough to make the stimulus
from which I is inferable worth processing. On the effect side, then, the
presumption is one of adequacy.
To achieve her communicative intention, the communicator has to
choose one of a range of different stimuli which would all make her
particular informative intention mutually manifest. We assume that she
eliminates any stimuli which would require too much effort on her part
(e.g. drawing a map when a verbal indication will do) or which she finds
objectionable (e.g. because of cultural rules prohibiting the use of certain
words). In most cases, this will still leave a wide range of possible stimuli.
It is in the interest of the addressee that the communicator should choose
the most relevant stimulus from that range: that is, the one that will call for
the least processing effort. Here the interests of communicator and
addressee coincide. Unless the communicator is merely pretending to
communicate, it is in her interest to be understood, and therefore to make
it as easy as possible for the addressee to understand her. An addressee
who doubts that the communicator has chosen the most relevant stimulus
compatible with her communicative and informative intentions - a hearer,
say, who believes that he is being addressed with deliberate and
unnecessary obscurity - might doubt that genuine communication was
intended, and might justifiably refuse to make the processing effort
required. All this is mutually manifest; it is therefore mutually manifest
that the communicator intends it to be manifest to the addressee that she
has chosen the most relevant stimulus capable of fulfilling her intentions.
On the effort side, then, the presumption is of more than mere adequacy.
The level of relevance that will be presumed to exist takes into account
158
Relevance
the interests of both communicator and audience. Let us call it a level of

optimal relevance. We can now spell out the presumption of optimal
relevance communicated by every act of.ostensive communication: c
(61) Presumption of optimal relevance

(a) The set of assumptions I which the communicator intends to
make manifest to tlle addressee is relevant enough to make it
worth the addressee's while to process the ostensive stimulus.
(b) The ostensive stimulus is the most relevant one the communicator
could have used to communicate I.
And here is the principle of relevance:
(62) Principle of relevance

Every act of ostensive communication communicates a presumption
of its own optimal relevance.
Let us now comment on the principle of relevance by raising and
answering a number of specific questions.
Does the principle of relevance apply to all forms of communication?

No: it applies only to ostensive communication, not to straightforward
coded communication. For instance, a telegraph employee who communicates messages by encoding them is expected to be accurate in her
encoding; she is not expected to produce particularly relevant stimuli.
To whom is the stimulus presumed to be relevant when there are no
definite addressees?
The addressees of an act of ostensive communication are the individuals
whose cognitive environment the communicator is trying to modify.
They can be specific individuals, as when Mary addresses Peter, or they
may be individuals falling under a certain description, as when we address
the present paragraph to all individuals who have read the book so far and
found it relevant to them. In broadcast communication, a stimulus can
even be addressed to whoever finds it relevant. The communicator is then
communicating her presumption of relevance to whoever is willing to
entertain it.
How reliable is the presumption of relevance?
As we all know, the world is full of bores. The principle of relevance
does not say that communicators necessarily produce optimally relevant
stimuli; it says that they necessarily intend the addressee to believe that
they do. Even bores manifestly intend their audience to believe that they
are worth listening to.
The presumption of relevance communicated by an utterance does not
have to be accepted as true. The communicator might fail to achieve
159
relevance; the addressee might doubt the communicator's ability to

succeed in being relevant. However, from the presumption of relevance
there follows a more reliable presumption that relevance has been
attempted, if not achieved. A communicator who fails to make it manifest
to her audience that she is being optimally relevant may nevertheless
succeed in making it manifest that she is trying to be optimally relevant.
However, ostensive communication must be seen as communicating more
than a mere presumption of attempted relevance. The addressee may be
willing to believe that the communicator has tried very hard to be relevant,
but if he also believes that she has totally failed, he will not pay attention to
her. So, however full of self-doubts she may be, a communicator must
intend to make it manifest to the addressee that her ostensive stimulus is
relevant enough.
Are you claiming that all ostensive communicators at least TR Y to be

optimally relevant?
This does not follow from the principle of relevance. Theoretically,
a communicator can communicate her presumption of relevance in bad
faith, just as she can communicate any assumption in bad faith. However,
it is generally true that ostensive communicators try to be optimally
relevant. When addressees are disappointed in their expectations of
relevance, they rarely consider as .a possible explanation that the
communicator is not really trying to be optimally relevant. It would be
tantamount to assuming that the apparent communicator is not really
addressing them, and perhaps not communicating at all. This rare
situation is illustrated by the case of filibusters.
Filibusterers make long speeches to an assembly merely in order to
delay its proceedings. All the usual features of verbal communication are
present and even salient, but for one: there js no attempt at optimal
relevance. Even if they tried, filibusterers could not hope to remain
relevant for the many hours, or even days, that a filibuster may last, and so
they do not keep their audience's attention, or even try to keep it. Are
filibusterers communicating, albeit defectively, or merely pretending to
communicate? For the apparent addressees, at least, it is clear that only a
pretence of communication is taking place, and that they are not being
genuinely addressed at all. It is like discovering that your host is putting in
front of you stuff whose edibility he has not even bothered to check. This
is tantamount to discovering that he is only pretending to feed you.
When no satisfactory level of relevance is achieved, a more plausible
assumption is that the communicator has tried to be optimally relevant,
but failed. Communicators take risks and sometimes fail, and addressees
expect such failures to occur occasionally. For example, if Mary knows
that Peter buys every book by Iris Murdoch, and she sees the latest one
160
Relevance
being put on display in the local bookshop, it would be reasonable for her
to say to Peter,
(63) Iris Murdoch's new book is in the bookshops.
It may turn out that Peter already has this information, in which case
utterance (63) will in fact be irrelevant to him. However, it would still have
been perfectly appropriate, and the presumption of relevance would have
been communicated in good faith, because Mary has at least tried to be
optimally relevant. Moreover, the risk she took was reasonable: it was
worth taking because of the hope, if she had succeeded, of achieving a high
degree of relevance to Peter.
How much effort the addressee can expect the communicator to put
into being relevant varies with the circumstances, the communicator, and
the relationship between communicator and addressee. Lecturers are
expected to try very hard to be relevant; students are allowed, and
sometimes even encouraged, to communicate without being hampered by
the fear of being irrelevant. A master talking to his servant may say
whatever he wishes and merely assume that it will be relevant enough; a
servant addressing his master is expected to have made quite sure that he
has something relevant to say.
How relevant is 'relevant enough to be worth the addressee's attention'?

We have assumed that an individual's cognitive resources are optimally
allocated when they yield the greatest cognitive effects. It might seem,
then, that to be worth the individual's attention, a stimulus must be more
relevant than any other external phenomenon, or internal representation,
that he could have been processing at the time. However, this does not
take the time factor into account.
Some phenomena and representations remain relevant and accessible
for a long time; others are both accessible and relevant only for a moment.
It is sometimes more efficient - that is, conducive to greater overall
relevance in the long run - to pay attention to a less relevant stimulus
whose cognitive effects might be lost forever if it is not immediately
processed, and to ignore some more relevant information which can as
well be processed later on. For instance, it may be consistent with the
principle of relevance to interrupt someone who is reading a fascinating
book in order to ask a mildly relevant but pressing question, or to draw his
attention to some moderately interesting incident in the landscape.
Similarly, some stimuli are of little intrinsic relevance but, by being
presented at the right time, increase the relevance of subsequent stimuli so
that a greater degree of overall relevance is achieved with them than
without them. This is generally true of the first sentence in a novel: though
of limited relevance in itself, it helps create a context in which subsequent
161
sentences will be more relevant. It is thus relevant enough to be worth the

reader's attention.
What counts as relevance enough, then, varies with the way in which
information is accessible, or can be made accessible, to the addressee over
time. It also varies with the degree of intellectual alertness of the addressee.
Imagine a group of people having a conversation in a cafe or a pub after
work, just a light conversation between friends. Here a modicum of
relevance should be enough: nobody will be willing to put in much
processing effort, or expect major contextual effects. For that matter,
nobody will put enormous effort into producing stimuli that would be
worth extensive processing. By contrast, consider what is supposed to
happen in a seminar. Here everyone is supposed to be on the alert, ready to
put a considerable amount of intellectual effort into producing and
processing information. In these circumstances, information relevant
enough to be worth the addressee's attention is quite relevant indeed.
There is little point, in one set of circumstances, in expecting a level of
relevance only normally achieved in quite different circumstances, and a
reasonable addressee will adjust his expectations accordingly.
The various factors we have mentioned are commonplace features of
everyone's everyday experience. It should not call for too much observation or imagination on the part of the communicator to estimate the
minimal level of relevance required. More specific considerations may
help. On various social occasions, the expected level of relevance is
culturally defined. In the course of a conversation, the level can be
adjusted, increased or decreased one step at a time. The addressee may
make manifest the minimal level of relevance he expects: by asking a
question, for instance. Even so, mistakes can occur. However, as we will
show, it is enough that the presumption of relevance should be communicated - and it always is. It does not have to be accepted as true in order to
fulfil its most important role: determining the interpretation of the
ostensive stimulus.
What are the differences between relevance theory and Grice's approach?
There are many. One is that the principle of relevance is much more
explicit than Grice's co-operative principle and maxims. Another is that
Grice assumes that communication involves a greater degree of cooperation than we do.
For us, the only purpose that a genuine communicator and a willing
audience necessarily have in common is to achieve uptake: that is, to have
the communicator's informative intention recognised by the audience.
Grice assumes that the communication must have 'a common purpose or
set of purposes, or at least a mutually accepted direction' (Grice 1975: 45)
over and above the aim of achieving uptake. We do not mean to deny that
162
Relevance
this is very often true, particularly in conversation. lO In a talk exchange,
a seminar or a book, there may well be a mutually manifest purpose or
direction. However, this does not follow from the principle of relevance,
and is not automatically conveyed by every ostensive stimulus. Knowledge of such a common purpose, when it exists, is one contextual factor
among others, and it is only as such that it can play a role in
comprehension. cl
Achieving optimal relevance, then, is less demanding than obeying the
Gricean maxims. In particular, it is possible to be optimally relevant
without being 'as informative as is required' by die current purposes of the
exchange (Grice's first maxim of quantity): for instance by keeping secret
something that it would be relevant to the audience to know. It seems to us
to be a matter of common experience that the degree of co-operation
described by Grice is not automatically expected of communicators.
Peo~le who don't give us all the information we wish they would, and
don't answer our questions as well as they could, are no doubt much to
blame, but not for violating principles of communication.
A more radical difference between Grice's approach and relevance
theory is this. Grice's principle and maxims are norms which communicators and audience must know in order to communicate adequately.
Communicators generally keep to the norms, but may also violate them to
achieve particular effects; and the audience uses its knowledge of the
norms in interpreting communicative behaviour.
The principle of relevance, by contrast, is a generalisation about
ostensive-inferential communication. Communicators and audience need
no more know the principle of relevance to communicate than they need
to know the principles of genetics to reproduce. Communicators do not
'follow' the principle of relevance; and they could not violate it even if
they wanted to. The principle of relevance applies without exception:
every act of ostensive communication communicates a presumption of
relevance. It is not the general principle, but the fact that a particular
presumption of relevance has been communicated by and about a
particular act of communication, that the audience uses in inferential
comprehension. ll
However, the most important difference between Grice's approach and
ours has to do with the explanation of communication. Grice's account of
conversation starts from a distinction between what is explicitly said and
what is implicated. No explanation of explicit communication is given;
essentially, the code model, with a code understood as a set of
conventions, is assumed to apply. Implicatures are explained as assumptions that the audience must make to preserve the idea that the speaker has
obeyed the maxims, or at least the co-operative principle. The principle of
163
relevance is intended to explain ostensive communication as whole, both

explicit and implicit. We will show in the next section how it does.
8 How relevance theory explains ostensive-inferential

communication
A communicator who produces an ostensive stimulus is trying to fulfil
two intentions: first, the informative intention, to make manifest to
her audience a set of assumptions I; and second, the communicative
intention, to make her informative intention mutually manifest. It is not
hard to see how the fulfilment of the communicative intention can lead to
the fulfilment of the informative intention: the realisation that a trustworthy communicator intends to make you believe something is an excellent
reason for believing it. This explains well enough why people engage in
ostensive communication. However, it does not explain how ostensive
communication works: how the communicative intention itself is fulfilled.
It is not obvious how the production of a stimulus can make the
communicator's informative intention mutually manifest, and thus lead to
the fulfilment of the communicative intention. As we have seen, with
other forms of intentional behaviour, evidence about the underlying
intentions is obtained by observing the effects of this behaviour. With
ostensive communication, the intended communicative effect is the
recognition of the informative intention. However, the intended informative effect does not generally occur, and thus cannot generally be
observed, until after the underlying informative intention has been
recognised. In that case, the informative intention cannot be inferred by
observing its independently achieved effects. The question is, how can it
be inferred at all?
Several inferential steps are needed if the informative intention is to
become mutually manifest. The stimulus has to make manifest, in the
mutual cognitive environment of communicator and audience, other
assumptions from which the informative intention can be inferred. First, it
must be manifest that the stimulus is ostensive. We have shown in section
6 how this can be achieved: by producing a manifestly intentional stimulus
which on the one hand attracts attention, and on the other is irrelevant
unless treated as evidence about the communicator's intentions. Once the
ostensive nature of a stimulus is mutually manifest to communicator and
addressee, it is also mutually manifest that the communicator has an
informative intention: that is, that she intends to make manifest to the
addressee some set of assumptions I. The problem of identifYing the
164
Relevance
communicator's informative intention reduces, then, to the problem of
identifying the set of assumptions I.
What the principle of relevance does is identify one member of I:
namely, the presumption of relevance. The presumption of relevance
is not just a member of I, it is also about I. As a result, it can be
confirmed or disconfirmed by the contents of I. The possibilities of
confirmation and disconfirmatIOn are different for the two different parts
(61a) and (61b) of the presumption of relevance, repeated here for
make manifest to the addressee is relevant enough to make it
worth the addressee's while to process the ostensive stimulus.
For the addressee, every assumption about the contents of I either
verifies (61a) - I is relevant enough - or else falsifies it. There may be
borderline cases, sets of assumptions on the margin of being relevant
enough. However, there cannot be cases for which there is insufficient
evidence, sets of assumptions whose relevance cannot be assessed by the
addressee: in processing I, he automatically discovers how relevant it is.
With the second part of the presumption of relevance, (61b), things need
not be so clear-cut. Given an assumption about the contents of I, it may
be manifest that the communicator could have used a more relevant
stimulus, and this will falsify (61b). However, (61b) may be neither
falsified nor verified: after all, in ordinary conditions, the addressee does
not know exactly what range of stimuli the communicator had at her
disposal, and hence cannot be sure that she has used the most relevant one
to communicate I. The presumption of relevance as a whole, then,
should either be clearly falsified (in the case where either (61a) or (61b) is
falsified), or be merely confirmed, but not verified (in the case where (61a)
is verified and (61b) is not falsified).
For some assumptions in I, all the evidence the communicator gives
the addressee is indirect: the addressee's only reason for accepting them is
the communicator's mutually manifest intention that he should. For other
assumptions in I, the communicator also provides direct evidence, as
when Peter ostensively leans back to let Mary see who is coming. The
status of the presumption of relevance is altered by the comprehension
process itself. At the start of the comprehension process, the initial
evidence for the presumption of relevance is entirely indirect; it is entirely
based on the communicator's guarantee that her stimulus is optimally
relevant to the addressee. However, by processing the stimulus, the
165
addressee naturally obtains direct evidence for or against the presumption

that it is optimally relevant; and by the end of the comprehension process
this direct evidence will have superseded the initial indirect evidence. In
intending to make the presumption of relevance manifest, the communicator must realise that she cannot help but provide direct and decisive
evidence for or against it. This is the crucial step towards identifying her
full informative intention, the set of assumptions I.
In trying to identify this informative intention, the addressee must
assume that the communicator is communicating rationally: that is, that
she has good reason to think that the stimulus she is producing will have
the intended effects. This applies not just to the identification of
informative intentions, but to the inferential identification of intentions in
general. Intentions are identified by assuming that the agent is rational,
and by trying to find a rational interpretation of her actions. It is not that
people in general, and communicators in particular, always suit their
means to their ends in a fully rational way. It is just that when they do not,
it is impossible to infer their intentions from their behaviour alone. In the
case of communicative behaviour, this compounds the irrationality, since
the success of communication depends on the addressee's ability to infer
the communicator's intentions.
A rational communicator, who intends to make the presumption of
relevance manifest to the addressee, must expect the processing of the
stimulus to confirm it. In other words, she must expect the contents of I
to verify (61a) and not to falsify (61b). To recognise the communicator's
informative intention, the addressee must discover for which set I the
communicator had reason to think that it would confirm the presumption
of relevance. We will argue that this is all he has to do.
The task of the addressee, then, is to construct possible interpretive
hypotheses about the contents of I, and to choose the right one. In
different circumstances and different cognitive domains, the task of
constructing and selecting a hypothesis may be carried out in different
ways. In some cases, it is best carried out by listing all the possible
hypotheses, comparing them, and choosing the best one. In others, it is
better carried out by searching for an initial hypothesis, testing it to see if it
meets some criterion, accepting it and stopping there if it does, and
otherwise repeating the process by searching for a second hypothesis, and
so on. To illustrate, suppose that Peter does not know exactly where he
left his sunglasses, but knows they are somewhere in the house. In one
case, he is away from home and has to telegraph Mary where to look for
the sunglasses. He should then make a mental list of all the places where he
might have left them, rank them in order of likelihood, and tell Mary the
most likely place. In another case, Peter is at home. He will take the first
hypothesis which occurs to him and look there; if he finds his sunglasses
166
Relevance
there, the search stops; otherwise, he will search a second place, and so on.
These two strategies, listing and ranking hypotheses, or searching for
hypotheses and testing them one by one, are suited to different types of
tasks. The first strategy is ill suited to tasks where it would be impossible
or inconvenient to list all possible hypotheses. For instance, if the task is to
find a pupil who is neither the tallest nor the shortest in the school, it
would be a waste of effort to rank all the pupils by height. The second
strategy is ill suited to tasks where there is no decisive criterion that can b
applied to isolated hypotheses. For instance, it would be impossible to
find out which is the tallest pupil in a school without taking all the pupils
mto account.
For other tasks, neither the list-and-rank strategy nor the item-by-item
testing strategy is appropriate on its own. The search for a true scientific
theory cannot be based on an examination of all possible theories, since we
do not know what these are; nor can it be based on a criterion which could
be used to decide whether an isolated theory is true. The strategy of
scientific discovery is much more complex, and involves both comparison
and individual testing; its results are, in principle at least, never final. As
we have pointed out, in this respect comprehension is unlike scientific
discovery: it yields final results almost immediately, which suggests that a
rather simple strategy must be involved.
Could comprehension be achieved by listing and ranking all possible
hypotheses about the communicator's informative intention? The idea
may seem attractive if comprehension is seen as a simple matter of
decoding a signal into a small set of possible messages and then choosing
among them. It must be rejected, however, because neither the possible
figurative interpretations of a coded message, nor its possible implicatures,
are enumerable. We will argue that this is true even when unambiguously
coded signals are used as stimuli. Moreover, even if it were possible to list
all the possible interpretations of an ostensive stimulus, it would still be
absurdly inconvenient. As we have seen, one of the factors which makes
one interpretation more relevant than others is that it requires less
processing effort. If the only way of finding the right interpretation were
to list and rank all possible interpretations, then all possible interpretations would require the same amount of effort: namely, the effort needed
to construct and compare them. It is hard to think of any ostensive
stimulus that would be worth such an absurd amount of effort.
Could comprehension be achieved, then, by constructing an initial
hypothesis, testing it, and moving on to a second one if the first is not
adequate? At first sight it might seem that here again the answer must be
no. Let us say that an interpretation is consistent with the principle of
relevance if and only if a rational communicator might have expected it to
be optimally relevant to the addressee.e Suppose now that the addressee
167
tests a possible interpretation and finds it consistent with the principle of

relevance. How could it follow that he would be right to select it? Surely
there are many other possible interpretations which might be consistent
with the principle of relevance too? Since consistency with the principle of
relevance is the only test we have, using it in an item-by-item testing
strategy will never yield the desired result.
This argument is mistaken. It misses the fact that the order in which
hypotheses are tested affects their relevance. As a result, the principle of
relevance does not generally warrant the selection of more than one
interpretation for a single ostensive stimulus. We will show that the
interpretation whose selection it warrants is the first one tested and found
to be consistent with the principle.
Consider first how an addressee who realises that an ostensive stimulus
has been produced, and hence that a presumption of relevance has been
communicated, might construct hypotheses about the communicator's
informative intention. First, the plausibility of some hypotheses may
already be manifest in the environment. Consider utterance (64):
(64) Peter (to Mary): Do you want some coffee?
By uttering (64), Peter makes it manifest that he wants an answer to his
question and that an appropriate answer would satisfy his expectations of
relevance. It is then plausible that the informative intention behind Mary's
next piece of communicative behaviour will be to make manifest an answer
to Peter's question.
The stimulus used by the communicator is itself a source of interpretive
hypotheses. The description of a non-coded ostensive stimulus (e.g. Mary
is sniffing ecstatically, or Mary is pretending to drive a car) gives immediate
access to the encyclopaedic entries of certain concepts and the assumption
schemas they contain. A coded stimulus gives immediate access to a highly
determinate set of concepts: the code itself determines which concepts are
activated, and moreover assembles them into a logical form which can be
directly used as an assumption schema. The context provides ways of
completing these assumption schemas into full hypotheses.
Once an initial set of hypotheses has been recovered, the addressee can
add to it by assuming that the set I includes further assumptions
contextually inferable from those already recovered. Moreover, by
extending the context, radically different hypotheses may become accessible. The important point is that, given the cognitive environment, given
the initial context, and given the stimulus, some hypotheses are more
accessible than others, and this means that they require less processing
effort.
Let us now reconsider the feasibility of the item-by-item testing
strategy. An addressee who is using this strategy, and who wants to
168
Relevance
maximise cognitive efficiency, will test hypotheses in order of accessibility. Suppose he arrives at a hypothesis which is consistent ~ith the
principle of relevance. Should he stop there, or go on and test the next
hypothesis on the ground that it might be consistent with the principle of
relevance too? It is easy to show that he should stop there. Suppose he
does go on, and finds another hypothesis which verifies the first part of the
presumption of relevance: the putative set I is relevant enough. In these
circumstances, the second part of the presumption of relevance is almost
invariably falsified. If it was at all possible, the communicator should have
used a stimulus which would have saved the addressee the effort of first
accessing two hypotheses consistent with the principle of relevance, and
then having to choose between them.
Consider the following utterance, for instance:
(65) George has a big cat.
In an ordinary situation, the first interpretation of (65) to occur to the

hearer will be that George has a big domestic cat. If it seems possible that
the speaker might have expected this interpretation to be optimally
relevant to the hearer, then he should stop there. Suppose he does not,
decides instead that the speaker might have expected other interpretations
to be optimally relevant too, and goes on searching for them. The word
'cat' is ambiguous: it may refer either to the domestic cat or to any animal
of the species Felis. Thus the hearer arrives at the hypothesis that (65)
might be intended to convey that George has a tiger, a lion, a jaguar, etc.
Maybe this information would be even more relevant than the fact that
George has a big domestic cat, thus verifying the first part of the
presumption of relevance. Nonetheless, the second part would automatically be falsified. A manifestly more relevant stimulus would have been
something-like (66), or, if the speaker lacked the necessary information,
something like (67) or (68):
(66) George has a tiger.
(67) George has a tiger or a lion, I'm not sure which.
(68) George has a felid.
These stimuli would have saved the addressee the effort of first accessing
and considering the 'domestic cat' interpretation, then accessing the 'felid'
interpretation, and then having to compare the two. Hence, the addressee
need not have bothered: the first interpretation consistent with the
principle of relevance was the best hypothesis. All other interpretations
would manifestly falsify the second part of the presumption of relevance.
When the communicator has an unbounded range of stimuli to choose
from, it follows from the second part of the presumption of relevance that
of all the interpretations of the stimulus which confirm the first part of
169
the presumption, it is the first interpretation to occur to the addressee

that is the one the communicator intended to convey. But what if the
communicator has a limited range of stimuli to choose from, so that if she
had intended to convey something other than the first optimally relevant
interpretation to occur to the addressee, she would have had no more
adequate stimulus at her disposal? In this case, either the first interpretation consistent with the principle of relevance is communicated, as before,
or nothing is communicated at all.
Imagine, for instance, a prisoner handcuffed and silenced. All she can
do before she is taken away is smile at her friend. How can he decide
whether she intended to convey something other than a sad goodbye, the
first plausible interpretation that occurs to him? And if she did, what did
she intend to convey? At first sight, there is no way of telling. However,
suppose that all the interpretations he can think of include a sad goodbye
as a subpart: then he can feel confident that this, at least, has been
communicated. By reasoning further, he should be able to see that the
prisoner herself can see that he has no way of crediting her with a fuller
informative intention, so that even though she may well have wished to
communicate more, she was not in a position rationally to intend to. So, at
most a sad goodbye has been communicated.
What if two essentially different interpretations seem to come simultaneously to the mind of the addressee, and they are both consistent with
the principle of relevance? In that case the addressee will be unable to
decide what the informative intention was, and communication will fail.
This is one of the few cases where an ambiguity is consciously perceived
during the comprehension process itself.
What if the communicator is mistaken in her presumption of relevance?
This will make the addressee's task a little more effort-consuming, and a
little more liable to failure, but not essentially different, and certainly not
impossible. To be consistent with the principle of relevance, an interpretation does not actually have to be optimally relevant to the addressee; it
must merely have seemed so to the communicator. Conversely, the first
optimally relevant interpretation may happen to be relevant in a way the
communicator could not have foreseen; in this case it is not consistent
with the principle of relevance. In every case, the task of the addressee is to
find an interpretation consistent with the principle of relevance - that is,
an interpretation which the communicator could manifestly have expected to be optimally relevant. This task is of course made easier, but not
essentially altered, when the addressee can trust the communicator, and
can therefore assume that the intended interpretation is actually the first
optimally relevant one to occur to him.
What happens when an unambiguously coded signal is used? Can the
criterion of consistency with the principle of relevance still be used ? Yes.
170
Relevance
Here, by way of illustration, is a political anecdote. In the Stalin era, two

friends in the West were arguing. Paul had decided to emigrate to Russia,
which he saw as a land of justice and freedom. He would go and write back
to Henry to let him know the beautiful truth. Henry tried to persuade him
not to go: there was oppression and misery in Russia, he claimed, goods
were scarce, and Paul's letters would be censored anyhow. Since Paul
would not be moved, Henry persuaded him to accept at least the following
convention: if Paul wrote back in black ink, Henry would know he was
sincere. If he wrote in purple ink, Henry would understand that Paul was
not free to report the truth. Six months after Paul's departure, Henry
received the following letter, written in black ink: 'Dear Henry, this is the
country of justice and freedom. It is a worker's paradise. In the shops you
can find everything you need, with the sole exception of purple ink .. .'
The point is that when a code is used in human communication, what
makes a communicated assumption manifest to the addressee is the
communicator's manifest intention to make it manifest. There is no way a
communicator could bind herself by a code or a convention to such an
extent that it would be impossible for her not to have the intention her
signal represents. The coded signal, even if it is unambiguous, is only a
piece of evidence about the communicator's intentions, and has to be used
inferentially and in a context. The hypothesis the signal suggests still has to
be tested for consistency with the principle of relevance, and if it fails to
meet this criterion, it must be rejected.
Contrary to first appearances, the principle of relevance does make it
possible to use an item-by-item testing strategy in comprehension. It
warrants the selection of the first accessible interpretation consistent with
the principle, if there is one, and otherwise no interpretation at all. In other
words, relevance theory explains how ostensive communication is
possible, and how it may fail.
Of course there are a lot of unanswered questions. For instance, how
exactly are assumption schemas filled out? What exactly determines the
order of accessibility of hypotheses? However, such questions are not
specific to relevance theory: they apply to cognitive psychology as a
whole. Since relevance theory is, among other things, an attempt to
ground models of human communication squarely in cognitive psychology, it cannot just take advantage of the insights of cognitive psychology,
but must also share its weaknesses. We have tried to show that the
relationship is not one way, and that relevance theory has contributions to
make to cognitive psychology. Other unanswered questions have more to
do with the study of communication proper, and verbal communication in
particular. What are the differences and relations between what is
explicitly communicated and what is implicitly communicated? How
does linguistic form affect interpretation? How are figurative interpreta-
171
tions determined? How is illocutionary force recognised? These questions

will be discussed in the next chapter.
4
Aspects of verbal
. .
communicatIon
In this chapter we want to outline some of the implications of relevance

theory for the study of verbal communication. What we are offering is
simply a sketch: we will not review the literature, we will discuss only
selected issues, and will not always justify our conclusions step by step.
However, we hope to show that relevance theory offers a pragmatic
framework in which serious questions can be raised, and new answers
developed.
1 Language and communication

Language and communication are often seen as two sides of a single coin.
On this view, the essential feature of language is that it is used in
communication, and the essential feature of communication is that it
involves the use of a language or code. The relation between language and
communication is thought of as like the relation between the heart and the
circulation of the blood: neither is properly describable without reference
to the other. In chapter 1, we argued that communication can be achieved
without the use of a code; in chapter 3, we showed how. In this section, we
want to complete the divorce between language and communication by
showing that languages, in a reasonably broad sense of t~le term, can and
do exist without being used for communication. Languages are indispensable not for communication, but for information processing; this is their
essential function. Having rejected the assumption that there is a necessary
link between language and communication, it then becomes interesting to
see what happens when, as a matter of contingent fact, they do become
linked: in verbal communication, for example.
In the broadest sense, a language is a set of well-formed formulas, a set
of permissible combinations of items from some vocabulary, generated by
a grammar. In a narrower sense, a: language is a set of semantically
Language and communication
173
interpreted well-formed formulas. A formula is semantically interpreted

by being put into systematic correspondence with other objects: for
example, with the formulas of another language, with states of the user of
the language, or with possible states of the world. A language in this
narrower sense - the one we will use - is a grammar-governed
representational system.
It would be possible to define a language even more restrictively: as a set
of semantically interpreted well-formed formulas used for communication.It would then be true by definition that language and communication
were inextricably linked. However, the definition itself would have to be
motivated. In science, a definition is motivated when it groups together
properties which are systematically linked in nature. Our point is
precisely that the property of being a grammar-governed representational
system and the property of being used for communication are not
systematically linked. They are found together in the odd case of human
natural languages, just as the property of being an olfactory organ and the
property of being a prehensile organ, though not systematically linked in
nature, happen to be found together in the odd case of the elephant's
trunk.
The activities which necessarily involve the use of a language (i.e. a
grammar-governed representational system) are not communicative but
cognitive. Language is an essential tool for the processing and memorising
of information. As such, it must exist not only in humans but also in a wide
variety of animals and machines with information-processing abilities.
Any organism or device with a memory must be able to represent past
states of the world or of itself. Any organism or device with the ability to
draw inferences must have a representational system whose formulas
stand in both syntactic and semantic relations to each other. Clearly, these
abilities are not confined to humans.
The great debate about whether humans are the only species to have
language is based on a misconception of the nature of language. The debate
is not really about whether other species than humans have languages, but
about whether they have languages which they use as mediums of
communication. Now the fact that humans have developed languages
which can be used to communicate is interesting, but it tells us nothing
about the essential nature of language. The originality of the human
species is precisely to have found this curious additional use for something
which many other species also possess, as the originality of elephants is to
have found that they can use their noses for the curious additional purpose
of picking things up. In both cases, the result has been that something
widely found in other species has undergone remarkable adaptation and
development because of the new uses it has been put to. However, it is as
strange for humans to conclude that the essential purpose of language is
174
Aspects of verbal communication
for communication as it would be for elephants to conclude that the

essential purpose of noses is for picking things up.
Language is not a necessary medium for communication: non-coded
communication exists. Nor is it necessarily a medium for communication:
languages exist which are not used for communication. However, language
is a ~ecessary attribute of communicating devices. Two devices capable of
communicating with each other must also be capable of internally
representing the information communicated, and must therefore have an
internal language. In the case of ostensive-inferential communication, this
internal language must be rich enough to repres"ent the intentions of other
organisms, and to allow for complex inferential processes.
In fact, for ostensive communication to be possible, the communicating
devices must have a richer internal language and more powerful inferential
abilities than are generally needed for coded communication. Bees do not
have to attribute intentions to one another or engage in inference in order
to communicate among themselves by means of their dance-based code:
all they need is an internal language capable of representing directions and
distances in space. Cognitively simple organisms can engage in coded
communication, whereas only cognitively sophisticated organisms can
engage in ostensive communication. Arguably, ostensive-inferential
communication exists within, and perhaps between, a variety of animal
species: for example, within those animal species which engage in
threatening behaviour and are able to distinguish threats from attacks;
perhaps between dog and human when the dog recognises its owner's
intentions.
It is clear that humans have an internal language rich enough for
ostensive-inferential communication. They also have external languages
such as Swahili or English, which are, of course, used for communication.
It might seem, then, that humans can communicate in two different ways:
either by ostension and inference, or by coding and decoding. We have
suggested a different view, which will be developed at length in this
chapter: that human intentional communication is never a mere matter of
coding and decoding. The fact is that human external languages do not
encode the kind of information that humans are interested in communicating. Linguistically encoded semantic representations are abstract
mental structures which must be inferentially enriched before they can be
taken to represent anything of interest.
Although the linguistic analysis of an utterance very much underdetermines its interpretation, the most striking feature of linguistic communication is that it can achieve a degree of precision and complexity
rarely achieved in non-verbal communication. When Mary sniffs ostensively to draw Peter's attention to the seaside smells, there is no limit to the
number of ways he can represent her behaviour to himself: there may be a
Language and communication
175
whole nebula of alternative interpretations, all closely similar in import

and comparable in relevance. To varying degrees, all non-verbal communication is weak communication in the sense defined in chapter 1: one
can never be sure which of a variety of assumptions made manifest by the
communicator she herself actually had in mind. The set of assumptions
which have been communicated can be defined in general terms, but the
individual members of this set cannot be enumerated.
With verbal communication, the situation is quite different. First, the
linguistic description of an utterance is determined by the grammar, and
does not vary with the interests or point of view of the hearers. Second,
this linguistic description yields a range of semantic representations, one
for every sense of the sentence uttered. Each semantic representation is a
schema, which must be completed and integrated into an assumption
about the speaker's informative intention, and can be as complex as the
speaker cares to make it. Moreover, each schematic sense is generally quite
different from all the others, and can be completed in quite different ways.
The various alternative interpretations of a non-coded ostensive stimulus,
of an appreciative sniff for instance, tend to form a continuous range of
variants; by contrast, the various possible interpretations of an utterance
tend to be radically different from one another, so that when one is
chosen, the others are automatically eliminated.
Consider utterance (1), for example:
(1) He's a bastard.
Let us assume that on the basis of a linguistic analysis of (1) and an
assignment of contextually accessible referents, the speaker might be
taken to be asserting any of (2a-d):
(2) (a) Peter is a nasty man.
(b) Bob is a nasty man.
(c) Peter is illegitimate.
(d) Bob is illegitimate.
It would be quite extraordinary for these various linguistically and
referentially possible interpretations of (1) to be equally consistent with
the principle of relevance. Because each alternative interpretation is
discrete and sharply distinguishable from the others, the hearer can
usually know for certain which one the speaker must have intended.
Linguistic communication is the strongest possible form of communication: it introduces an element of explicitness where non-verbal communication can never be more than implicit. Of the assumptions conveyed
by an utterance, at least those that are explicitly conveyed can be
enumerated.
We regard verbal communication, then, as involving two types of
176
communication process: one based on coding and decoding, the other on

ostension and inference. The coded communication process is not
autonomous: it is subservient to the inferential process. The inferential
process is autonomous: it functions in essentially the same way whether or
not combined with coded communication (though in the absence of coded
communication, performances are generally poorer). The coded communication is of course linguistic: acoustic (or graphic) signals are used to
communicate semantic representations. The semantic representations
recovered by decoding are useful only as a source of hypotheses and
evidence for the second communication protess, the inferential one.
Inferential communication involves the application, not of specialpurpose decoding rules, but of general-purpose inference rules, which
apply to any conceptually represented information.
Incidentally, this view of verbal communication has implications about
the origin of human languages. The fact that the semantic representations
of natural-language expressions are merely tools for inferential communication suggests that inferential communication had to exist before
external languages developed: human external languages are of adaptive
value only for a species already deeply involved in inferential communication. Remember the old comparison between language and money: words
and currencies are similar in that they both derive their value from
convention? We would like to push the comparison in a different
direction. Money is central to a modern, monetary economy, just as
language is central to verbal communication. However, the monetary
system could only appear in a pre-existing economic system, and only
makes sense as part of such a system. Similarly, human natural language
could only appear in a pre-existing inferential communication system, and
it only makes sense as part of such a system. Verbal communication is a
specifically human enhancement of ostensive-inferential communication.
2 Verbal communication, explicatures and implicatures

An utterance is a perceptible modification of the physical environment. As
such, it makes manifest a variety of assumptions. Suppose, for instance,
that Mary utters the complex sound transcribed in (3):
(3) [It~getbuld]
This makes manifest to Peter a set of assumptions A which might include,
among many others, assumptions (4a-e):
(4) (a) Someone has made a sound.
(b) There is someone in the house.
Explicatures and implicatures
177
(c) Mary is at home.

(d) Mary has spoken.
(e) Mary has a sore throat.
If some of (4a-e) are relevant to Peter, then Mary's behaviour may be

relevant simply by making these assumptions manifest to him. In that
case, the linguistic, and in particular the semantic properties of the
sentence uttered make no contribution to relevance. A clearing of the
throat might have been relevant to Peter in just the same way; or rather, it
would have been more relevant, since it would have achieved the same
effects without needing any linguistic processing at all.
The set A of assumptions made manifest by Mary's behaviour also
includes (5):
(5) Mary has uttered the sentence 'It will get cold.'
In appropriate conditions, an assumption of the form in (5) will be
automatically constructed. Even in poor acoustic conditions, a phonetic
stimulus in the hearer's native language is automatically analysed as a
token of a particular linguistic structure: [It~getbuld] is analysed as 'It
will get cold.' This information may be filtered out sub-attentively, but as
long as minimum standards of acoustic clarity and salience are met, the
phonetic signal will be automatically analysed and assigned a semantic
representation (or, in the case of ambiguity, several semantic representations), making manifest an assumption of the form in (5).
In other words, a linguistic stimulus triggers an automatic process of
decoding. Just as we cannot choose to see the objects around us in black
and white rather than in colour, just as we cannot choose not to hear a gun
going off nearby, so we cannot choose to hear an utterance in a language
we know as merely an unanalysed stream of sounds. We automatically
recover its semantic representation, even if we accidentally overhear it and
know it was not meant for us, or even (as the evidence on binaural
shadowing shows)! if we are not conscious of hearing it at all. The
linguistic decoding system has all the hallmarks of automatic, reflex
perceptual systems such as hearing and vision. In the terms of Fodor
(1933), who develops this point at length, it is an input system rather than a
central processing system, and this is one reason why it has been so
relatively amenable to study. This suggests in turn that if comprehension
is defined as a process of identifying the speaker's informative intention,
linguistic decoding is not so much a part of the comprehension process as
something that precedes the real work of understanding, something that
merely provides an input to the main part of the comprehension process.
Verbal communication is never achieved merely by the automatic
decoding of linguistic signals. Such decoding occurs even when it is
178
Aspects of verbal communication.
manifest that no attempt at communication is being made: for instance

when an actor doing voice exercises is accidentally overheard. It also occurs
when an utterance is used to communicate information which bears no
relation to its semantic content, as in the following dialogue:
(6) A:
B:
A:
B:
Did your treatment for stammering work?

Peter Piper picked a peck of pickled pepper.
How amazing!
Yes, b-b-but th-th-that's not s-s-something I v-v-very often
w-w-want to s-s-say.
B's first reply communicates that the treatment worked well, not by
saying so, but by producing direct evidence that it did. Properly speaking,
this is not a case of verbal communication, and it falls outside the scope of
pragmatics. Verbal communication proper begins when an utterance, such
as B's second reply, is manifestly chosen by the speaker for its semantic
propertIes.
In other words, verbal communication proper begins when the speaker
is recognised not just as talking, not even just as communicating by
talking, but as saying something to someone. Most utterances do this, of
course, and an adequate account of verbal communication must explain
why. One way of explaining it is to assume that people learn, or are
innately equipped with, more or less ad hoc pragmatic rules to the effect
that utterances should be used for communication only in virtue of their
semantic properties. 2 However, this leaves exceptions such as (6) to be
explained.
A simpler explanation follows from the principle of relevance. According to relevance theory, the correct interpretation of an ostensive stimulus
is the first accessible interpretation consistent with the principle of
relevance. For most utterances, this will be an interpretation based on
semantic properties: the other properties of the utterance are generally not
relevant enough to yield an interpretation consistent with the principle of
relevance. In odd cases such as (6), the semantic properties of the
utterance do not yield an appropriate interpretation and other properties
(in this case acoustic properties) do. The principle of relevance thus
explains both the usual, semantically based cases of utterance interpretation and the occasional exceptions.
Suppose that Mary's behaviour is an ordinary case of verbal communication - that is, that it makes manifest assumption (7):
(7) Mary has said to Peter 'It will get cold.'
Since saying something to someone is a case of ostensive communication,

the set A of assumptions made manifest by Mary's utterance includes
(8):
179
(8) There is a set of assumptions I which Mary intends to make manifest

to Peter by saying to him 'It will get cold.'
The task of the hearer can now be described in two ways. One is to say
that the hearer must find in A a mutually manifest assumption of the form
in (9):
(9) The speaker intends to make I manifest.
However, suppose that the speaker achieves not only her communicative
intention but also her informative intention - as she will if the hearer both
understands her and trusts her enough. Then I, the set of assumptions
communicated by the utterance, will be a subset of A, the set of
assumptions made manifest by the utterance. The hearer's task can then be
described in another way: the hearer must decide which assumptions in
A would also, if the speaker were trustworthy, be members of I: that
is, he must decide which assumptions made manifest by her utterance are
such that it is mutually manifest that the speaker intended to make them
manifest.
Typically, the set I might include assumptions such as (lOa-e):
(10) (a) Mary's utterance is optimally relevant to Peter.
(b) Mary has said that the dinner will get cold very soon.
(c) Mary believes that the dinner will get cold very soon.
(d) The dinner will get cold very soon.
(e) Mary wants Peter to come and eat dinner at once.
The goal of pragmatics is to explain how the hearer's task, as described
above, can be carried out: how he can identify a set I, e.g. (10), using as
premises a description of the speaker's behaviour, e.g. (7), together with
contextual information.
The hearer's task involves a variety of inferential sub-tasks. The first is
to assign the utterance a unique propositional form. This involves
disambiguating the sentence uttered: that is, selecting one of the semantic
representations assigned to it by the grammar. Here, a single sense of
'cold' (experiencing cold or inducing cold) must be selected. However, the
recovery of a unique propositional form involves more than disambiguation. A referent must be assigned to each referring expression (e.g. 'It' in
our example). The contribution of vague terms such as 'will' must be made
more specific (e.g. by the addition of very soon in our example). In other
words, a semantic representation must be selected, completed and
enriched in various ways to yield the propositional form expressed by the
utterance. This task is inferential- that much is uncontroversial. However
there is very little in the pragmatic literature to explain how it is carried out
- apart from the comment that Gricean maxims and mutual knowledge
180
might help.3 Moreover, the complexity of the task is generally underestimated: it is seen as simply a matter of choosing a single sense and reference
from a limited set of alternatives. The fact that logical forms must often be
enriched is generally ignored; no explanation is given of how such
enrichment can be achieved.
Suppose Peter has decided that 'It' refers to the dinner, that 'will' refers
to the immediate future, and that 'cold' means inducing cold. In other
words, he has decided that the propositional form expressed by Mary's
utterance is (lad):
(10) (d) The dinner will get cold very soon.
An utterance does more than express an explicit propositional form: it

expresses this form in a certain linguistically determined mood. For
instance, if Mary's utterance (3) has a falling intonation contour, it will be
in a declarative mood: it will be a case of 'saying that'. If it has a rising
intonation contour, it will be in an interrogative mood: it will be a case of
'asking whether'. Mood is linguistically encoded, but just as the logical
form of an utterance underdetermines the propositional form expressed,
so the mood of an utterance underdetermines the propositional attitude
expressed. One of the hearer's sub-tasks, again an inferential one, is to
identify this propositional attitude.
Having identified the propositional form of an utterance and the mood
expressed, the hearer is in a position to identify one further member of I
(apart, that is, from the presumption of relevance itself): namely, the
assumption that the speaker has expressed this particular propositional
form in this particular mood. For instance, suppose that Mary's utterance
is in a declarative mood. Then it is mutually manifest that Mary intended
(lab) to be manifest to Peter; in other words, it is inferable that (lab) is a
member of I:
(10) (b) Mary has said that the dinner will get cold very soon.
However, a hearer can recover (lab) but still not know what propositional
attitude Mary intended to communicate; and without knowing this, he
will be unable to decide what she intended to communicate apart from
(lab) itself. In particular, even though Mary has said that the dinner will
get cold very soon, she need not be asserting that the dinner will get cold
very soon. Asserting that P involves communicating that one believes that
P. However, in the weak sense of 'saying that' which corresponds to the
declarative mood, one can say that P without communicating that one
believes that P. For example, in saying that the dinner will get cold very
soon, Mary might be speaking metaphorically or ironically, in which case
she would not communicate that she believes that the dinner will get cold
very soon.
181
Moreover, asserting that P involves more than merely communicating

that one believes that P. Hence, Mary might communicate that she
believes the dinner will get cold very soon, without actually asserting it.
Suppose that (lOc) is a member of I:
(10) (c) Mary believes that the dinner will get cold very soon.
A speaker who communicates that she believes that P does not automatically communicate that P. For instance, suppose it is mutually manifest
that Peter believes that the dinner will stay hot for as long as it takes him to
finish what he is doing, and that he has no reason to trust Mary's opinion
here more than his own. Then Mary could not have intended her utterance
to achieve relevance by making manifest to Peter that the dinner would get
cold very soon, but only by making manifest that she believes it will.
We will discuss problems of figurative interpretation and illocutionary
force in sections 7-10. Let us suppose, for the moment, that Peter has
decided that Mary intended to communicate both that she believes that the
dinner will get cold very soon, and that the dinner will get cold very soon.
In other words, let us suppose that it is mutually manifest that Mary
intended Peter to infer (lOd) from (lOc):
(10) (d) The dinner will get cold very soon.
An utterance which meets this condition, i.e. which communicates its
propositional form, we will call an ordinary assertion.
Suppose now that from (lOd), together with mutually manifest
information, (lOe) is inferable:
(10) (e) Mary wants Peter to come and eat dinner at once.
Suppose, moreover, that it is mutually manifest that it is contextual
implication (lOe) which makes the whole utterance relevant enough to be
worth Peter's while to process. Then it is inferable that (lOe) is a member
of I, and (lOe) is communicated by Mary's utterance.
However, there is a striking difference between the way (lOb-d) on the
one hand, and (lOe) on the other, are identified. Assumptions (lOb-d)
include as sub-parts one of the logical forms encoded by the utterance.
They are constructed inferentially, by using contextual information to
complete and enrich this logical form into a propositional form, which is
then optionally embedded into an assumption schema typically expressing an attitude to it. Let us call this process of assumption construction the
development of a logical form. (lOe), by contrast, is not a development of
one of the logical forms encoded by the utterance; it is constructed on the
basis of contextual information, and in particular by developing assumption schemas retrieved from encyclopaedic memory. For instance, Peter's
182
encyclopaedic memory might contain a whole scenario of 'dinner at

home', including the assumption schema (11):
(11) Mary wants Peter to come and eat dinner at time t. [t = time at which
dinner is still hot]
We see the difference between (lab-d) on the one hand and (10e) on the
other as a difference between explicit and implicit communication. We
define:
(12) Explicitness a
An assumption communicated by an utterance U is explicit if and
only if it is a development of a logical form encoded by U.
On the analogy of 'implicature', we will call an explicitly communicated
assumption an explicature. Any assumption communicated, but not
explicitly so, is implicitly communicated: it is an implicature. By this
definition, ostensive stimuli which do not encode logical forms will, of
course, only have implicatures.
This classificatory concept of explicitness lends itself quite naturally to a
comparative interpretation. An explicature is a combination of linguistically encoded and contextually inferred conceptual features. The smaller
the relative contribution of the contextual features, the more explicit the
explicature will be, and inversely. Explicitness, so understood, is both
classificatory and comparative: a communicated assumption is either an
explicature or an implicature, but an explicature is explicit to a greater or
lesser degree.
This is an unconventional way of drawing the distinction between the
explicit and implicit 'content' of an utterance. On a more traditional view,
the explicit content of an utterance is a set of decoded assumptions, and the
implicit content a set of inferred assumptions. Since we are claiming that
no assumption is simply decoded, and that the recovery of any assumption
requires an element of inference, we deny that the distinction between the
explicit and the implicit can be drawn in this way.
Grice sees things rather differently. For him, recovering the explicit
content of an utterance apparently amounts to recovering what we would
call the propositional form and mood expressed; any other assumptions
communicated by the utterance, whether decoded or inferred, is an implicature. Decoded implicatures are what he calls 'conventional implicatures'; inferred implicatures are 'non-conventional', the most familiar of
these being the famous 'conversational implicatures'. We would deny that
there are conventional implicatures in Grice's sense, but this is not our
main reservation about his way of drawing the distinction between the
explicit and the implicit.
The identification of propositional form
183
The main problem with Grice's distinction has to do not with the
characterisation of implicatures, but with the characterisation of the
explicit. First, he does not envisage the kind of enrichment of logical form
involved, for instance, in interpreting 'will' as will very soon; he treats
comparable cases, for instance the interpretation of 'and' as and then in
some contexts, as cases of implicature. Most Gricean pragmatists assume
without question that any pragmatically determined aspect of utterance
interpretation apart from disambiguation and reference assignment is
necessarily an implicature. In fact, recent work has shown that a number
of problems with classical implicature analyses are resolved when the
'implicatures' are re analysed as pragmatically determined aspects of
explicit content. 4
Second, Grice says very little about how propositional attitudes are
communicated, and it is unclear what he would regard as 'explicit' and
what 'implicit' here. Third, he has no notion of degrees of explicitness.
Generally speaking, we see the explicit side of communication as richer,
more inferential, and hence more worthy of pragmatic investigation than
do most pragmatists in the Gricean tradition.
In the next two sections, we show how relevance theory accounts for
the recovery of the propositional form of an utterance (section 3) and its
implicatures (section 4)~ For simplicity of exposition, we will look only at
ordinary assertions, i.e. utterances which communicate their propositional forms. In the final sections of this chapter we will generalise our
treatment to other types of utterance.
3 The identification of propositional form

The hearer's first task in recovering the explicatures of an utterance is to
identify its propositional form. In this section, we will describe this task in
more detail and show how it is carried out. We will restrict our attention to
ordinary assertions, in which the propositional form of the utterance is
itself an explicature.
The task is, of course, to identify the right propositional form, and the
right propositional form is the one intended by the speaker. However, this
cannot be the criterion the hearer uses to identify the right propositional
form: if he already knew the speaker's intention, he would have no task of
identification left. What criterion does the hearer use to select the right
propositional form? Although there is a considerable literature on
disambiguation and reference assignment, this question has not been
seriously addressed. Experimental studies of disambiguation simply take
for granted that there is normally only a single sense of an utterance which
184
looks as if it could have been intended by the speaker; no attempt is made

to explain why this is so. The aims of psycho linguists working on
disambiguation lie elsewhere: they want to describe not the criterion used
in disambiguation, but the procedure by which disambiguation is
achieved.
Could the answer simply be that the right propositional form is the one
obtained by going through a certain procedure (just as the right result in
multiplication is the one obtained by applying a certain algorithm)? The
existence of so-called garden-path utterances is strong evidence that this is
not so. Consider (13), with the possible interpretations (14a-b):5
(13) I saw that gasoline can explode.
(14) (a) I saw that it is possible for gasoline to explode.
(b) I saw that can of gasoline explode.
When (13) is processed in isolation, the normal disambiguation procedure
favours interpretation (14a). However, the continuation in (15) would
force a reinterpretation:
(15) And a brand new gasoline can it was too.
What such garden-path utterances strongly suggest is that the outcome of

the normal disambiguation procedure is not automatically accepted as the
right propositional form. It is rejected if it fails to meet some criterion
which has yet to be defined.
At the end of chapter 3, we made a suggestion about what the general
criterion for the interpretation of an ostensive stimulus might be: the right
interpretation is the one that is consistent with the principle of relevance.
This in turn suggests a criterion for identifying the propositional form of
an utterance: the right propositional form is the one that leads to an overall
interpretation which is consistent with the principle of relevance. Let us
say that in this case the propositional form itself is consistent with the
principle of relevance.
Whatever regular procedures there are for disambiguation, reference
assignment and enrichment, they yield at best a tentative identification of
propositional form, an identification which will be rejected if it turns out
not to be consistent with the principle of relevance. This is why
interpretation (14a) is rejected: the explicatures recovered by regular
disambiguation procedures from (13), the first part of the overall
utterance, do not lead to an interpretation which is consistent with the
principle of relevance once (15), the second part of the utterance, is taken
into account.
Our suggestion is, then, that the propositional form the hearer should
be interested in recovering is the one that is consistent with the principle of
relevance. The next question is, what general procedure might the hearer
185
use to identify propositional forms which meet this criterion? Here again,
the outline of an answer is strongly suggested by the principle of
relevance. At every stage in disambiguation, reference assignment and
enrichment, the hearer should choose the solution involving the least
effort, and should abandon this solution only if it fails to yield an
interpretation consistent with the principle of relevance.
We will now look separately at the three sub-tasks involved in the
identification of propositional form: disambiguation, reference assignment and enrichment. One problem we immediately encounter is that we
cannot avoid the use of artificial examples. When an artificial example is
produced, say as part of a theoretical discussion or in an experimental
situation, it is processed and understood in isolation from any natural
context. This is not to say that it is processed and understood in isolation
from any context. In the first place, it gives access to encyclopaedic
information about the objects and events referred to, and hence to a range
of potential contexts of the usual type; in the second place, the author or
experimenter may provide some elements of a natural context by
describing a setting, asking the individual to imagine a previous utterance,
and so on.
Even so, artificial examples tend to favour considerations of effort over
considerations of effect in the assessment of relevance. In the absence of
real-life contextual constraints, or constraints specially set up by the
experimenter, hearers automatically construct a context which yields the
least effort-consuming conceivable interpretation. It would thus be easy,
on the basis of artificial examples, to conclude that the identification of
propositional form is entirely determined by a principle of least effort.
The existence of garden-path utterances such as (13) should prevent us
from making such a mistake.
Though effort is only one of the two factors involved in the assessment
of relevance, it is a factor well worth studying, and here there is an
advantage in the fact that it is to some extent isolated by artificial examples.
We are assuming that the identification of propositional form involves
two mental mechanisms: a linguistic input module and a central inferential
ability. How are the two mechanisms related, and how does the effort
made by each affect the overall processing effort? More specifically, does
the linguistic input module construct all the possible semantic representations of a sentence, one of which is then selected by centtal processes? Or
are the semantic representations of a sentence more or less effortconsuming for the input module to construct, so that the easiest one is
constructed first, a second representation being constructed only if the
first is rejected, and so on? In other words, how are the 'wrong'
interpretations filtered out?
These questions are not going to be answered at the purely speculative
186
level of the present discussion. The considerable experimental evidence

already available is not conclusive, judging by the debates it has
provoked. 6 What can be said at the speculative level is this: the
filtering-out of all interpretations but one is not conscious, which strongly
suggests that it is a relatively peripheral process. On the other hand, what
makes one interpretation 'right' and another 'wrong' is contextual
information: mainly general encyclopaedic information in the case of
artificial examples.
Consider (16), for instance:
(16) The child left the straw in the glass.
This could mean either that the child left the drinking tube in the glass, or
that the child left the cereal stalks in the glass. In the absence of a special
context, it is the drinking-tube interpretation which is selected. Why? At a
purely linguistic level, there is no reason to assume that the cereal-stalk
sense of 'straw' is less accessible than the drinking-tube sense; no reason,
then, why one interpretation should be preferred. The selection manifestly involves contextual factors.
A child drinking from a glass with a straw is a stereotypical event which
we assume, as do most other people working on the organisation of
memory, is recorded in the form of a single chunk, stored at a single
location in memory and accessed as a single unit. Such a chunk constitutes
a highly accessible encyclopaedic context in which the drinking-tube
interpretation of (16) can be processed at minimal cost. There is nothing,
of course, to prevent a child leaving a bunch of cereal stalks in a glass, or a
speaker choosing to report such an event. However, the encyclopaedic
context needed to process this information would be less accessible than
the context needed to process the drinking-tube interpretation of (16): it
would not be stored as a chunk, but would have to be derived by collecting
together information about children and glasses on the one hand, and
cereal stalks on the other. Hence the more easily accessible drinking-tube
interpretation of (16), once recovered, is also more easily processed.
If we assume, with Fodor (1983), that input modules have no access to
general encyclopaedic information, examples such as (16) seem to imply
that the input module has to construct all the semantic representations of
an utterance, the wrong ones then being filtered out at a central level after
all. However, the relationship between input module and central processes need not be that simple: for instance, the input module might
construct all the linguistically possible interpretations of the first constituent of the sentence, and submit them to the central mechanism, which
would, when possible, choose one of them and inform the linguistic
module of its choice. As a result, the module's decoding processes would
be partly inhibited; it would retain only those interpretations of the next
187
constituent which are linguistically compatible with the selected interpretation of the first constituent, and so on. With the interaction of input
module and central mechanisms so conceived, it remains true that ~the
module has no access to encyclopaedic contextual information; however,
contextual factors may affect its processes in a purely inhibitory way.
For example, when (17), the first part of (13), has been decoded, the
central mechanism is in a position to choose between an interpretation on
which 'that' is a demonstrative determiner, and one on which it is a
complementiser:
(17) I saw that ...
(13) I saw that gasoline can explode.
Demonstrative determiners need a particular type of context: one created

by pointing, for instance. In an artificial situation, the complementiser
interpretation, which does not need an ad hoc context, is less effortconsuming and will be preferred. Assuming that from then on the
operations of the input module are restricted accordingly, interpretations
of (13) on which 'that' is a demonstrative will be automatically filtered out
at the modular level, and 'can' will be interpreted as a verb, not a noun.
Disambiguation hypotheses are recovered by decoding and evaluated
inferentially. Hypotheses about the intended reference of referring
expressions are not generally recoverable by decoding alone. 7 To construct a hypothesis about the reference of 'It' in (18), the hearer must use
not only linguistic but also non-linguistic information:
(18) It will get cold.
Linguistically, the only constraint on the reference of 'It' is that it should

not refer to a human. This leaves the hearer an indefinitely large choice of
referents.
How should the hearer construct and evaluate referential hypotheses?
Given the principle of relevance, b he should first consider the immediate
context, see if any of the concepts of a non-human entity represented in
this context, when substituted for 'It', yields a propositional form
consistent with the principle of relevance; if not, he should extend the
context and repeat the procedure. This may sound like a cumbersome
performance, but in practice it can be quite simple. Suppose the hearer
knows that dinner is on the table, and is wondering whether it will stay hot
long enough for him to finish a letter he is writing: then all the contextual
implications of (18) will already have been worked out, and they only need
strengthening to yield an immediately accessible chunk of contextual
effects. In this case, the hearer has no difficulty in testing the dinner as a
possible referent for 'It', or in checking that the resulting overall
interpretation is consistent with the principle of relevance. It is in such a
188
situation that (18) would be most appropriate. If the immediate context

does not yield an adequate referent for 'It', the hearer might add to the
context the encyclopaedic entries of the various concepts which have cold
as their lexical entry. One highly accessible schema these entries would
yield is about meals getting cold. The relevance of the resulting
interpretation could be easily checked.
It is widely believed that if all but one of the senses of a sentence are
eliminated and referents are assigned to its referential expressions, the
resulting combination of sense and reference corresponds to a unique
propositional form. We have argued againsHhis view. Consider (19):
(19) The bat is grey.
Suppose that 'bat' is understood in the zoological sense, that 'The bat'
refers to a specific bat, and that the present tense of 'is' refers to a specific
time. Then it is standardly claimed that (19) is capable of being true or
false; it expresses a unique propositional form. Maybe so, but what about
(20)-(22)?
(20) Peter's bat is grey.
(21) The bat is too grey.
(22) The bat is big.
'Peter's bat' might refer to the bat owned by Peter, the bat chosen by
Peter, the bat killed by Peter, the bat mentioned by Peter, and so on
indefinitely. It is hard to believe that the genitive is ambiguous, with as
many senses as there are types of relationship it may be used to denote, or
that all these relationships fall under a single definition which is the only
meaning expressed by use of the genitive on any given occasion. It seems,
rather, that the semantic interpretation of a sentence with a genitive from
which ambiguities and referential indeterminacies have been eliminated is
still something less than fully propositional. Contextual information is
needed to resolve what should be seen as the semantic incompleteness,
rather than the ambiguity, of the genitive.
It can be similarly argued that an adverb such as 'too' is semantically
incomplete. A bat is too grey for something. If you do not know what that
something is, you do not fully know what 'too grey' is being used to
express. (21) is a perfectly grammatical sentence of English. Yet a
combination of one of its senses with fixed references corresponds to an
indefinite range of propositional forms. Again similar arguments applyand have often been applied-to scalar adjectives such as 'big' in (22): is the
bat big for an adult bat, big for a bat of its age, big for a pet, etc.? And does
'big' without a scale of reference express a complete meaning?
Examples such as (20)-(22) strongly suggest that the gap between
semantic representations and propositional forms cannot be closed merely
189
by disambiguation and reference assignment. Quite often, semantlC
representations must also be enriched. This task is, of course, an
inferential one. Consider (23):
(23) It will take some time to repair your watch.

The interpretation recoverable from this utterance by decoding and
reference assignment is a truism and thus irrelevant. It goes without saying
that watch-repairing is a process with a temporal duration, and a speaker
aiming at optimal relevance must have intended to express something more than goes without saying. In general, an utterance of the
form in (23) should be interpreted as conveying not the truism that the job
in question will take some time, but that it will take an amount of time it
would be relevant to remark on: i.e. longer than would otherwise be
expected. Suppose I always take my watch to the same watchmaker, and it
usually takes about a week to repair. Then if the speaker of (23) is aware of
these facts, she must be understood as saying that the repair will take
longer than a week. The more precise the expectations, the more precisely
the speaker's intentions can be pinned down.
This situation is predicted by relevance theory in the usual way: an
utterance, like any other ostensive stimulus, is a piece of evidence about
the communicator's informative intention. The fact that it activates certain
concepts and, in the case of utterances, a certain logical form, is ground for
assuming that at least some of the assumptions which the communicator
intends to make manifest contain these concepts or this logical form. The
logical form of an utterance, in particular, is an assumption schema. The
presence of semantically incomplete or manifestly vague terms is a clear
indication of where the schema might be enriched. In the case of 'some
time' in (23), it is a matter of finding the first accessible enrichment of the
concept which will yield an interpretation relevant enough to be
consistent with the principle of relevance. The 'some time' in question
might be at least one second, at least one hour, at least one week, and so on,
each of these interpretations being an enrichment of the preceding one in
the sense that it contains the same information and more. In this case, the
first accessible enrichment consistent with the principle of relevance is the
one which specifies that the time it will take to repair the watch is at least
more than would normally be expected.
Similarly, compare (24) and (25):
(24) I have had breakfast.
(25) I have been to Tibet.
What can be recovered from these utterances by decoding and reference
assignment is that the speaker has had breakfast, or been to Tibet, at some
point within a period of time preceding her utterance. In real life, a hearer
190
would be expected to make some more or less specific assumption about

how long that period was. In this, he is guided by the fact that a
presumption of relevance has been communicated. In the case of (24), for
example, it would normally go without saying that the speaker had had
breakfast at some point in her life. If she intends her utterance to be
manifestly relevant, she must intend to make manifest that she has had
breakfast recently enough for it to be worth remarking on: for example,
recently enough not to be in immediate need of food. In the case of (25), by
contrast, the mere fact that the speaker had visited Tibet at some point in
her life could well be relevant enough, and in'the absence of more specific
information this is the interpretation that would be consistent with the
principle of relevance.
Let us show informally how disambiguation, reference assignment and
enrichment combine, by looking at how (26) might be interpreted as a
continuation of (27) and (28):
(26) Peter's bat is too grey.
(27) Your team is disqualified from the baseball game.
(28) We have chosen John's mouse for our breeding experiment.
Suppose (26) is a continuation of (27) in a real-life situation in which the

hearer is a member of a baseball team. (27) gives him access to his
encyclopaedic entries for baseball games, teams, including his own
baseball team, and disqualification. It is also likely to raise in his mind the
question of why his team has been disqualified. Suppose his team has a
member called Peter, who has been playing with a particular grey baseball
bat. In the circumstances, he could scarcely avoid the hypothesis that the
speaker has said that the baseball bat his team-mate Peter has been playing
with is too grey to be used in a regulation baseball game. This hypothesis
would be retroactively strengthened by yielding an adequate range of
contextual effects in an easily accessible context: in particular, by
explaining why his team has been disqualified. It is this interpretation that
would be consistent with the principle of relevance.
Suppose (26) is a continuation of (28) in an artificial situation, say in a
disambiguation experiment. (28) gives the hearer access to his encyclopaedic entries for choosing, for mice, for breeding and for experiments; to
achieve any degree of relevance, however, he would have to make some
assumption about who the speaker was, who John was, what john's
relation to the mouse was, and what the breeding experiment was for. If he
has a schema for classroom biology experiments, he might have easy
access, for example, to the assumption that the speaker is a schoolteacher,
that John is a schoolboy, and that he has brought the mouse as a possible
subject for a classroom experiment on genetics. The same schema can be
191
reused, at minimal processing cost, to interpret the second part of the

utterance, yielding the hypothesis that Peter is another schoolboy who has
brought a grey bat as a possible subject for the same classroom experiment
on genetics, but that this bat is too grey to be used in this particular
experiment. This hypothesis would again be retroactively strengthened by
yielding a satisfactory range of contextual effects in this stereotypical
context. In particular, suppose that the interpretation of (28) has raised in
the hearer's mind the question of why that particular mouse was chosen
for that particular experiment; then (26), on this interpretation, would
provide an answer to this question. The principle of relevance thus plays a
decisive role in the recovery of the propositional form of the utterance,
and therefore of its explicatures, in artificial situations as in natural ones.
The above discussion, apart from outlining our particular hypotheses
about disambiguation, reference assignment and enrichment, raises a
more general question about the role of semantic representations in
communication. There is a widespread view that all the thoughts that a
human might entertain and want to communicate could in principle be
linguistically encoded. Katz incorporates this view in the following
'principle of effability':
(29) Each proposition (thought) is expressible by some sentence in every
natural language. (Katz 1981: 226)
What does it mean to say that every thought is expressible by some

sentence? On a weak interpretation, it means that every thought can be
conveyed by uttering some sentence. If no limit is placed on the
complexity of the sentence, this seems a matter of common sense. It is this
commonsense intuition which is the strongest and most obvious argument
for the effability principle. However, on this interpretation, the claim
made by the effability principle is about utterances in context rather than
about sentences: about language use rather than about language in itself. It
does not entail that every entertainable thought can be linguistically
encoded.
Katz offers a stronger and more interesting interpretation of the
principle. According to him, for every thinkable thought there is, in every
language, a sentence one of whose senses uniquely corresponds to that
thought; if that sentence is used literally and in that sense, then, whatever
the context, it expresses that thought. According to this view, every
thought is encoded by a sense of some sentence.
On this view it would be possible, at least in principle, to communicate
thoughts linguistically without any appeal to inference and context
(except, perhaps, for purposes of disambiguation). Why, then, do natural
languages contain so many sentences which encode not thoughts but
192

merely incomplete logical forms? Why is it that most sentences actually
uttered are schematic, requiring inference as well as decoding for full
comprehension? Katz's main answer is that
it allows speakers to make use of contextual features to speak far
more concisely than otherwise. Imagine how lengthy utterances
would be if everything we wanted to express had to be spelled out
explicitly in the grammar of our sentences. Pragmatics saves us from
this wasteful verbosity. Thus, instead of using sentences like [(30)],
we can, on occasion, use sentences like [(31)].
"
[(30)] The man who just asked the stupid question about the relation
between the mental and the physical has, thank God, left the
room.
[(31)] Thank God, he is gone. (Katz 1977: 19-20)
Note, however, that (30) is not fully propositional: it would express
different propositional forms in different situations by referring to
different individuals. Its interpretation might need fewer contextual clues
than that of (31), but it would need some. To eliminate the referential
indeterminacy, what would be needed is something like (32), where time
and space could be specified in terms of universal co-ordinates:
(32) Thank God, the man x who at time t was in location l has, at time t' ,
left the room which the man x was in at time t.
However, it is open to questIon whether either (30) or (32) expresses the

same thought as (31). That is, I may think what is conveyed by 'Thank
God, he is gone' without entertaining any of the senses of (30) or (32) or of
any other such sentence; I need not describe to myself the man whose
departure I am rejoicing over as 'the man who just asked the stupid
question about the relation between the mental and the physical' or as 'the
man who at time t was in location l', or in terms of any external-language
definite description. It seems plausible that in our internal language we
often fix time and space references not in terms of universal co-ordinates,
but in terms of a private logbook and an ego-centred map; furthermore,
most kinds of reference - to people or events for instance - can be fixed in
terms of these private time and space co-ordinates. Thoughts which
contain such private references could not be encoded in natural languages
but could only be incompletely represented.
What does this imply for the possibility of two people having exactly
the same thought, and for the possibility of communication? It implies
that two people may be able to think of the same man that he has gone,
without being able to think exactly the same thought, because they might
not individuate him in exactly the same way. Similarly, by saying 'He has
193
gone' I may induce in you a thought which is similar to mine in that it

predicates the same thing (that he is gone) of the same individual, but
which differs from mine in the way you fix the reference of 'He'. It seems
to us neither para40xical nor counterintuitive to say that there are
thoughts that we cannot exactly share, and that communication can be
successful without resulting in an exact duplication of thoughts in
communicator and audience. We see communication as a matter of
enlarging mutual cognitive environments, not of duplicating thoughts. 8
If sentences do not encode thoughts, what do they encode? What are the
meanings of sentences? Sentence meanings are sets of semantic representations, as many semantic representations as there are ways in which the
sentence is ambiguous. Semantic representations are incomplete logical
forms, i.e. at best fragmentary representations of thoughts. We have
argued that they are incomplete in more than one way: not just because
they contain indeterminate referring expressions such as pronouns, but
also because they contain underdefined constituents such as 'too', 'some
time', or the genitive. What we are suggesting is that the claim that the
semantics of natural languages might be too weak to encode all humanly
thinkable thoughts is quite compatible with what is known of the role of
language in verbal communication.
One entertains thoughts; one does not entertain semantic representations of sentences. Semantic representations of sentences are mental
objects that never surface to consciousness. If they did, they would seem
entirely uninteresting (except, of course, to semanticists). Semantic
representations become mentally represented as a result of an automatic
and unconscious process of linguistic decoding. They can then be used as
assumption schemas to identify first the propositional form and then the
explicatures of an utterance. It is these explicatures alone that have
contextual effects, and are therefore worthy of conscious attention.
4 The identification of implicatures
In the last section we showed how the principle of relevance guides the
identification of propositional form. In section 10, we will discuss the
identification of the speaker's propositional attitude. From the context,
the propositional form of the utterance and the propositional attitude
expressed, all the explicatures of the utterance can be inferred. For the time
being, we will continue to look only at ordinary assertions, where the
propositional form is itself an explicature, and indeed the explicature on
which most of the contextual effects of the utterance, and therefore most
of its relevance, depend. In this section, we will show how the principle of
194
relevance guides the recovery of implicatures. We will argue that the

implicatures of an utterance are recovered by reference to the speaker's
manifest expectations about how her utterance should achieve optimal
relevance.
A speaker may have reason to believe that certain information would be
relevant to her hearer, without having the faintest idea what its relevance
will be. A passer-by asks you the time; you know it is 5 p.m. The fact that
he has asked this question gives you reason to believe that the information
that it is 5 p.m. will be relevant to him. However, you have no way of
knowing how it will be relevant: in what corttext it will be processed and
what its contextual effects will be. Intuitively, in this situation, the simple
answer that it is 5 p.m. will carry no implicatures at all. Your informative
intention in giving this answer would merely be to make manifest that it is
5 p.m. This is the first inferable interpretation consistent with the principle
of relevance.
Contrast this with a case where the speaker does have manifest
expectations about how her utterance will be relevant:
(33) (a) Peter: Would you drive a Mercedes?
(b) Mary: I wouldn't drive ANY expensive car.
We assume that (33b) is an ordinary assertion, and hence that its main
explicature, the only one we will be concerned with, is simply its
propositional form. The propositional form of (33 b) does not directly
answer the question in (33a). However, it gives Peter immediate access to
his encyclopaedic information about expensive cars, which includes, let us
suppose, the information in (34):
(34) A Mercedes is an expensive car.
If processed in a context containing assumption (34), (33b) would yield

the contextual implication (35):
(35) Mary wouldn't drive a Mercedes.
This should, in turn, have an acceptable range of contextual effects in a
context which Peter, by asking whether Mary would drive a Mercedes, has
indicated that he has accessible.
We have a situation, then, in which Mary, in producing (33b), has not
directly - i.e. explicitly - answered Peter's question, but has made
manifest a contextually implied answer. Given that, in normal circumstances, she could not expect her utterance to be relevant unless it made
manifest such an answer, it is mutually manifest that this implied answer is
intentional: it is an implicature of her utterance. An implicature is a
contextual assumption or implication which a speaker, intending ~er
utterance to be manifestly relevant, manifestly intended to make manifest
195
to the hearer. We will distinguish two kinds of implicatures: implicated

premises and implicated conclusions. (34) is an implicated premise of (33b),
and (35) an implicated conclusion. All implicatures, we claim, fall into one
or the other of these two categories.
Implicated premises must be supplied by the hearer, who must either
retrieve them from memory or construct them by developing assumption
schemas retrieved from memory. What makes it possible to identify such
premises as implicatures is that they lead to an interpretation consistent
with the principle of relevance, and that they are manifestly the most easily
accessible premises to do so. Implicated conclusions are deduced from the
explicatures of the utterance and the context. What makes it possible to
identify such conclusions as implicatures is that the speaker must have
expected the hearer to derive them, or some of them, given that she
intended her utterance to be manifestly relevant to the hearer. Thus,
implicated premises and conclusions are both identifiable as parts of the
first inferable interpretation consistent with the principle of relevance.
Implicatures (34) and (35) have two properties which many pragmatists
think of as shared by all - or at least all theoretically interesting implicatures. In the first place, they are fully determinate. Mary expects
Peter to supply not merely something like premise (34) and conclusion
(35), but a premise and conclusion with just this logical content. Second,
Mary is entirely responsible for their truth. Suppose that before (33b) was
produced, Peter had mistakenly thought that Mercedes cars were cheap;
then (33b) would provide as much disconfirmation of this assumption as if
Mary had explicitly asserted that a Mercedes was an expensive car. Or
suppose Peter had merely suspected that Mercedes cars were expensive;
then (33b) would strengthen this assumption as much as if Mary had
explicitly asserted that a Mercedes was an expensive car. In other words,
Mary is just as responsible for the truth of (34) and (35) as if she had
asserted them directly.
There has been a tendency in modern pragmatics to treat all implicatures along these lines: as fully determinate assumptions for which the
speaker is just as much responsible as if she had asserted them directly. On
this approach, utterance comprehension consists in the recovery of an
enumerable set of assumptions, some explicitly expressed, others implicitly conveyed, but all individually intended by the speaker.
Grice himself does not regard implicatures as determinate:
Since, to calculate a conversational implicature is to calculate what
has to be supposed in order to preserve the supposition that the
Cooperative Principle has been observed, and since there may be
various possible specific explanations, a list of which may be open,
the conversational implicatum in such cases will be a disjunction of
196
such specific explanations; and if the list of these is open, the

implicatum will have just the kind of indeterminacy that many actual
implicata do in fact seem to possess. (Grice 1975: 58)
Other pragmatists/ while recognising the existence of indeterminacy,
have tended to exclude it from consideration. Thus, Gazdar comments:
because indeterminacy is hard to handle formally, I shall mostly
ignore it in the discussion that follows. A fuller treatment of
implicature would not be guilty of this omission, which is only really
defensible on formal grounds. (Gazdar 1979: 40)
The proposal to ignore indeterminacy might be seen as a legitimate
idealisation, a simplifying assumption of the kind which would pass
unquestioned in other domains of scientific inquiry and should need no
justification here. It is reasonable, the argument goes, to look first not at
the complex, fuzzy reality which we know exists, but at an idealisation
from which the fuzziness has been eliminated, and which is amenable to
formal treatment. If the implicatures of an utterance are treated as a
determinate set of intended inferences, an explicit theoretical model can be
set up, which can later be filled out in various ways to account for the
fuzziness of the full range of data.
However, not every idealisation is legitimate. An idealisation is not
legitimate if, in simplifying the data, it introduces some significant
distortion which puts theoretical work on the wrong track. An example of
such an illegitimate idealisation is the reduction of a language by
pre-Chomskyan linguists to a finite corpus of utterances. We will argue
that by concentrating on fully determinate implicatures such as (34) and
(35) above, modern pragmatists have obscured an important difference
between explicit content and implicit import. As a result, they have
perpetuated a mistaken semiotic view of communication and in particular
have deprived themselves of the ability to provide an adequate analysis of
stylistic and poetic effects.
Notice that although, in producing (33b) above, Mary manifestly
expects Peter to derive the conclusion in (35) and all the implications of
(35) he might be interested in, if this is all she expects, she cannot assume
that her utterance is optimally relevant. If the entire relevance of (33b)
depends on the recovery of (35), Mary could have spared Peter some
unnecessary processing effort by saying (36) instead:
(36) I wouldn't drive a Mercedes.
It follows from the principle of relevance that in giving the indirect answer
in (33b), she must have expected to achieve some additional contextual
197
effects not obtainable from (36), which would offset the additional effort
needed to process (33b), supply premise (34) and deduce (35) as an
implicated conclusion. More generally, it follows from the principle of
relevance that the surplus of information given in an indirect answer must
achieve some relevance in its own right.
It does not follow, though, that there is any specific implicature, apart
from (34) and (35), which Mary must have expected Peter to recover. An
act of communication merely makes manifest which assumptions the
communicator intends to make manifest, or, equivalently, it merely makes
these assumptions manifest on the further assumption that the communicator is trustworthy. It does not necessarily make the audience actually
entertain all the assumptions communicated. This is true of implicatures
too. Implicatures are merely made manifest by the act of communication
(again, on the further assumption that the speaker is trustworthy). Some
implicatures are made so strongly manifest that the hearer can scarcely
avoid recovering them. Others are made less strongly manifest. It is
enough that the hearer should pay attention to some of these weaker
implicatures for the relevance of the intended interpretation to become
manifest.
As we have seen, utterance (33b) gives Peter access to his encyclopaedic
information about expensive cars. One 'obvious line of interpretation
would be to retrieve the names of other expensive cars, and derive the
conclusion that Mary would not drive them. It is a stereotypical - and
hence highly accessible - item of general knowledge that a Rolls Royce
and a Cadillac are expensive cars. Hence it would be reasonable for Peter
to add premises (37) and (38) to the context, derive conclusions (39) and
(40), and investigate their contextual effects:
(37)
(38)
(39)
(40)
A Rolls Royce is an expensive car.

A Cadillac is an expensive car.
Mary wouldn't drive a Rolls Royce.
Mary wouldn't drive a Cadillac.
Or he could construct some premise such as (41), which is plausible

enough in their mutual cognitive environment, derive conclusion (42) and
investigate the contextual effects of this conclusion:
(41) People who refuse to drive expensive cars disapprove of displays of
wealth.
(42) Mary disapproves of displays of wealth.
The indirect answer in (33 b) thus opens up a number of possibilities of
interpretation not available for its direct counterpart (36). Given the
principle of relevance, Mary must have expected some of these possibilities to be fruitful enough to offset the extra processing costs incurred.
198
Are (37)-(42) implicatures of (33b)? Not under the idealisation

described above. In the first place, Mary need not have specifically
intended Peter to supply premises (37), (38) and (41) and derive
conclusions (39), (40) and (42). (33b) has a number of different contextual
implications in an appropriately extended context, any of which might
yield enough contextual effects to offset the extra processing effort
involved. In the second place, precisely because different subsets of
implicated premises and conclusions might be used to establish the
optimal relevance of the intended interpretation, none of them need have
been specifically intended by Mary. Mary's mutually manifest intention is
merely to make manifest some such assumptions. Hence she does not
make any of these assumptions more than weakly manifest. She does not
guarantee their truth as strongly as she guarantees the truth of (34) and
(35). Thus, whereas by producing (33b) Mary provides conclusive
evidence that she regards a Mercedes as an expensive car and would refuse
to drive in one, she provides rather less than conclusive evidence that she
would refuse to drive in a Rolls Royce.
On the other hand, it would be nonsense to say that by producing (33b)
Mary has not encouraged Peter to think she would refuse to drive in a
Rolls Royce. Short of explicitly asserting it, or of actually forcing him to
supply it as an implicature, what clearer encouragement could she have
given him than (33b)? Although (39), (40) and (42) cannot be forced into
the mould of fully determinate, specifically intended inferences, it would
be clearly wrong to regard them as entirely unintended, as derived on
Peter's sole responsibility. As we have seen, Mary would not have been
justified in communicating her presumption of relevance if she had not
expected some of these implicatures to be derived - if she had not
intended, therefore, to make all of them weakly manifest.
Let us pursue this line of argument by considering some other premises
and conclusions that Peter might be tempted to supply in processing the
indirect answer (33b). In an ordinary modern cognitive environment, it is
manifest that if Mary regards a Mercedes as expensive she will also regard a
Rolls Royce and a Cadillac as expensive, and hence that (37)-(40) are fair
reflections of her views. It is also manifest that she will regard as expensive
any other car which costs as much as or more than a Mercedes. But which
are these? Relatively safe premises and conclusions such as (37)-(42) shade
off into those, such as (43)-(46), with rather greater risks attached:
(43)
(44)
(45)
(46)
An Alfa Romeo is an expensive car.

A BMW is an expensive car.
Mary wouldn't drive an Alfa Romeo.
Mary wouldn't drive a BMW.
Are these implicatures of (33b)? While in no way forcing Peter to
199
investigate these possibilities, Mary has certainly given him some

encouragement to think along these lines, although the conclusions
derived must be treated with rather more caution than either the fully
determinate implicatures (34) and (35) or the strongly invited inferences
(37)-(42).
Imagine now that Peter believes (47) and finds it worth his effort to
derive (48) as a contextual implication from (47) and the explicature of
(33b):
(47) People who would not drive an expensive car would not go on a cruise
either.
(48) Mary would not go on a cruise.
It is very doubtful that Mary has given Peter any encouragement to supply
the premise in (47) and derive the conclusion in (48). What examples
(33 )-( 48) show is that there may be no cut-off point between assumptions
strongly backed by the speaker, and assumptions derived from the
utterance but on the hearer's sole responsibility. The fiction that there is a
clear-cut distinction between wholly determinate, specifically intended
inferences and indeterminate, wholly unintended inferences cannot be
maintained. Relevance theory offers a way of getting rid of this fiction
without sacrificing clarity of conceptual framework.
Let us say that the implicatures of an utterance - like assumptions in
general- may vary in their strength. To communicate an assumption A i~
to make mutually manifest one's intention to make A manifest or mort
manifest. The greater the mutual manifestness of the informative intention
to make manifest some particular assumption, the more strongly thi~
assumption is communicated. The strongest possible implicatures an
those fully determinate premises or conclusions, such as (34) and (35)
which must actually be supplied if the interpretation is to be consistent with
the principle of relevance, and for which the speaker takes full responsibility.
Strong implicatures are those premises and conclusions, such as (37)-( 42),
which the hearer is strongly encouraged but not actually forced to supply. Tht
weaker the encouragement, and the wider the range of possibilities among
which the hearer can choose, the weaker the implicatures. Eventually, as wa~
illustrated with (47)-( 48), a point is reached at which the hearer receives no
encouragement at all to supply any particular premise and conclusion, and hE
takes the entire responsibility for supplying them himself.
On this approach, the indeterminacy of implicatures presents nc
particular formal problem. An utterance with a fully determinatt
implicated premise or conclusion forces the hearer to supply just thi~
premise or conclusion and attribute it to the speaker as part of her beliefs
An utterance ,,ith a small range of strongly implicated premises 01
200
conclusions strongly encourages the hearer to use some subset of these
premises or conclusions, and to regard some subset of them - not
necessarily the same subset - as part of the speaker's beliefs. An utterance
with a wide range of weakly implicated premises or conclusions again
encourages the hearer to use some subset of these assumptions, and to
regard some subset of them - again not necessarily the same - as part of the
speaker's beliefs. Clearly, the weaker the implicatures, the less confidence
the hearer can have that the particular premises or conclusions he supplies
will reflect the speaker's thoughts, and this is where the indeterminacy
lies. However, people may entertain different thoughts and come to have
different beliefs on the basis of the same cognitive environment. The aim
of communication in general is to increase the mutuality of cognitive
environments rather than guarantee an impossible duplication of
thoughts.
To conclude this section, we want to contrast our approach with other
approaches to implicature. First, in our framework, there is no connection
between conveying an implicature and violating a pragmatic principle or
maxim. Gricean implicatures fall into two classes: those where there is no
violation or where the violation is only apparent, and those where the
violation is genuine and even the recovery of an implicature does not
restore the assumption that the maxims have been observed. For us, this
second class of examples must be reanalysed.
In the second place, we have taken seriously Grice's requirement that
implicatures should be calculable: that is, recoverable by an inference
process. In Grice's framework, and the framework of most pragmatists,
some sort of ex post facto justification for the identification of an
implicature can be given, but the argument could have worked equally
well for quite different assumptions which happen not to be implicated at
all. This is particularly true of the second class of implicatures, those
resulting from deliberate violation of the maxims; they tend to violate the
calculability requirement in a particularly blatant way.
Consider Grice's analysis of irony, for example. Mary says (49), and in
doing so patently violates the first maxim of quality (truthfulness):
(49) Jim is a fine friend.
Peter, assuming that Mary must have been trying to convey some true
information, looks around for some true assumption related to (49),
which she might have wanted to convey. He decides that she must have
wanted to convey the opposite of what she has said:
(50) Jim is not a fine friend.
Hence, on Grice's analysis a speaker may deliberately violate the maxim of

truthfulness and succeed in implicating the opposite of what she has said.
201
However, 'looking around for some related assumption which the speaker
might have wanted to convey' does not count as an inference process:
such a process is virtually free of rational constraints. Why, for instance,
should Peter not decide that (49) is to be interpreted as conveying the
closely related (51), as long as (51) is something that Mary might have
wanted to convey?10
(51) Bill is a fine friend.
Relevance theory does not sanction the analysis of (49) as meaning (50)
- unless (49) is recognisably a slip of the tongue - if only because a speaker
who merely wanted to convey (50) could have spared her hearer some
unnecessary processing effort by asserting it directly. In section 9 we will
propose an alternative account.
The reason why standard accounts of implicature do not always satisfy
the calculability requirement on implicatures is that the calculation of
implicatures is a matter of non-demonstrative inference. It involves a
partly non-logical process of assumption formation; then the assumption
has to be confirmed. Standard accounts impose few if any constraints on
the formation of assumptions. In practice, they just take what is the
intuitively correct assumption and show that it is consistent with Gricean
maxims or with some other principles, constraints or rules of the same
kind. Intuitively wrong assumptions, such as the assumption that the
speaker of (49) means (51) are, alas, just as easy to 'confirm' in this way.
Relevance theory solves this problem by looking not just at the
cognitive effects of an assumption, but also at the processing effort it
requires. The psychological processes by which assumptions are formed
determine their accessibility, which affects their relevance, which affects
their plausibility. Different assumptions are thus predicted to differ in
plausibility before any confirmation process takes place. When an initially
more plausible interpretation is found to be consistent with the principle
of relevance, then it is uniquely confirmed, and all less initially plausible
interpretations are disconfirmed.
Another important point to have emerged from this section has to do
with the scope of pragmatics. The idea that pragmatics should be
concerned purely with the recovery of an enumerable set of assumptions,
some explicitly expressed, others implicitly conveyed, but all individually
intended by the speaker, seems to us to be a mistake. We have argued that
there is a continuum of cases, from implicatures which the hearer was
specifically intended to recover to implicatures which were merely
intended to be made manifest, and to further modifications of the mutual
cognitive environment of speaker and hearer that the speaker only
intended in the sense that she intended her utterance to be relevant, and
hence to have rich and not entirely foreseeable cognitive effects. Pragmat-
202
ists and semioticians who look only at the strongest forms of implicature
have a badly distorted image of verbal communication. They miss, or at
least fail to explain, the subtler effects achieved by much implicit
communication. We will return to the role of weak implicatures when we
discuss style and tropes in sections 6 and 7. But first we want to consider
some of the ways in which linguistic form affects pragmatic interpretation.
5 Propositional form and style: presuppositional effects

A speaker who intends to produce a relevant utterance has two related
aims: first, to create some contextual effect in the hearer, and second, to
minimise the processing effort this involves. It might seem that two
utterances with the same linguistically determined truth conditions must
have identical contextual effects. We will show that, on the contrary, they
may differ both in their contextual effects and in the processing effort they
require, and that this is the key to an explanatory theory of style.
In this section we will look at a range of stylistic effects essentially
determined by the linguistic structure of the utterance and achieved in the
very process of identifying its propositional form. Since, as before, we are
looking only at ordinary assertions, this propositional form is also the
main explicature of the utterance. These stylistic effects have been
approached in terms of various distinctions: between topic and comment,
given and new, theme and rheme, presupposition and focus, presupposition and assertion, and so on, and are illustrated by the following sets of
examples: 11
(52)(a)
(b)
(53)(a)
(b)
(c)
(54)(a)
(b)
(c)
Bill's twin sister lives in BERLIN.

Bill has a twin sister who lives in BERLIN.
It rained on MONDAY.
On Monday it RAINED.
On MONDAY it rained.
John-Paul the Second is the present POPE.
The present Pope is John-Paul the SECOND.
It is John-Paul the SECOND who is the present Pope.
In (52a), the information that Bill has a twin sister is presupposed, or

treated as given; in (52b), it is asserted, or treated as new. (53a)-(53c)
illustrate a variety of effects obtainable by stress placement and wordorder change: thus (53a) or (53c), unlike (53b), would be an appropriate
answer to the question 'When did it rain?', whereas (53 b), unlike (53a) or
(53c), would be an appropriate answer to the question 'What was the
weather like on Monday?' In (54a) the topic is intuitively John-Paul the
Second, whereas in (54b) and (54c) it is the present Pope.
Propositional form and style
203
There is a huge descriptive literature in this area, but nothing

approaching an explanatory theory of the relation between linguistic
structure and pragmatic effects. 12 There are, however, a number of
scattered insights which seem to us worth pursuing. One is that it is
natural for given information (i.e. information the speaker is treating as
known or uncontroversial) to come before new and for focal stress to fall
towards the end of the utterance, since this in some way facilitates
comprehension. This is to a certain extent common sense, but it is not
always true that given information comes before new: new information
comes before given in our examples (53c) and (54c) above, and Green
(1980) surveys a wide range of falsifying examples. The problem is to
produce a theory that accommodates both the 'natural' and the 'marked'
examples.
Another idea is that stress is a sort of vocal equivalent of pointing, a
natural means of drawing attention to one particular constituent in an
utterance. The parallel is reinforced by the fact that stress, like pointing, is
inherently ambivalent. In (55), for example, the stressed noun 'FOOTBALL' is part of the noun phrase 'the football match', the verb phrases 'to
see the football match' and 'went off to see the football match', and the
sentence 'Susan went off to see the football match':
(55) Susan went off to see the FOOTBALL match.
As is well known, a speaker who puts focal stress on 'football' may be
intending to highlight anyone of these more inclusive syntactic constituents. Let us call the smallest stressed constituent, in this example the
noun 'football', the focally stressed constituent, and the constituent that it
is used to highlight the focus. Then the focally stressed constituent rarely
determines a unique focus, and the problem is to show how an actual focus
is chosen from a range of potential foci.
A third idea is that the way to find the focus of a declarative utterance is
to see what Wh-question it was designed or could be appropriately used
to answer. For example, (55) can be interpreted as an answer to a series of
related questions, each the result of substituting an appropriate Whphrase for one of its possible foci: 'Which match did Sus an go off to see?',
'What did Sus an go off to see?', 'What did Susan go off to do?', 'What did
Susan do?', and 'What happened?' Each possible focus determines a
Wh-question, and vice versa. Although this intuitive observation is made
by virtually everyone working in the area, the problem is to provide some
satisfactory explanation of why it should be so.
Finally, it is often suggested that instead of binary distinctions between
given and new, focus and presupposition and so on, what is really involved
is a gradient or hierarchical structure. Suppose that the focus of (55) above
is the verb phrase 'to see a FOOTBALL match', so that all the information
204
carried by this verb phrase is highlighted. Nonetheless it is intuitively clear

that it is not all equally highlighted: that the information carried by the
word 'football' is more prominent than the information carried by the
word 'see'. It is as if the focus consisted of a nested series of foci with
varying degrees of prominence, the smallest focus being the most
prominent of all. Again, the intuitions are clear, although it is less clear
how they should be incorporated into an explicit theory.
What gives these observations explanatory value is the idea that the
syntactic and phonological organisation of an utterance may directly
affect the way it is processed and understood. What is puzzling is that
having seen the possibility of a natural linkage between linguistic form and
pragmatic interpretation, so many authors feel the need to interpose
intermediate levels of semantic and pragmatic description to link artificially what, if these insights are correct, is already naturally linked. We would
like to pursue the idea of a natural linkage between linguistic form and
pragmatic interpretation, and show how it might be worked out within
the framework of relevance theory.
Consider first what in our framework would be the most uneconomical
way of processing an utterance. If processing costs were no object, the
hearer could explore all possible parsings, disambiguations, illocutionary
forces, reference assignments and enrichments. He could take each
resulting explicature, extend the immediate context by adding to it the
encyclopaedic entries of all the concepts appearing in the explicature, and
systematically explore its contextual effects in that context. He could,
moreover, derive all the analytic implications of the explicatures, add to
the context the encyclopaedic entries of all their constituent concepts,
explore the resulting set of contextual implications, and so on indefinitely.
This method of processing would guarantee that no conceivable interpretation would be overlooked, no possible context left unexplored, and
no possible contextual effect left underived. Clearly, however, it would
also involve a lot of fruitless processing.
Notice, though, that because an utterance is produced and processed
over time, the hearer will be in a position to access some of its constituent
concepts, with their associated logical and encyclopaedic entries, before
others. For a speaker aiming at optimal relevance, efficient exploitation of
this temporal sequencing will be crucial. Here we will show briefly how it
might help to 'hold down the costs of disambiguation and reference
aSSignment.
The sooner disambiguation and reference assignment are achieved, the
less processing effort will be required. The greater the number of possible
interpretations that have to be borne in mind as the utterance proceeds, the
greater the processing effort. It follows that a speaker aiming at optimal
relevance should phrase her utterance so as to facilitate early - and correct
- disambiguation. How might this be achieved?
205
Much recent work on parsing suggests that parsing is to some extent a

'top-down' process: that the hearer constructs anticipatory hypotheses
about the overall structure of the utterance on the basis of what he has
already heard. 13 For example, he might not only identify each word and
tentatively assign it to a syntactic category, but use his knowledge of its
lexical properties and syntactic co-occurrence restrictions to predict the
syntactic categories of following words or phrases.
The experimental literature on disambiguation suggests that disambiguation and reference assignment are also to some extent 'top-down'
processes: that the hearer makes anticipatory hypotheses about the overall
logical structure of the utterance and resolves potential ambiguities and
ambivalences on the basis of these. 14 We want to suggest a way of
constructing anticipatory logical hypotheses on the basis of the anticipatory syntactic hypotheses whose role in comprehension seems fairly well
established.
Let us assume that logical forms, like syntactic forms, are trees of
labelled nodes (or, equivalently, labelled bracketings). The syntactic labels
are the categories N, NP, V, VP, and so on, where N might be regarded as a
variable over nouns, NP as a variable over noun phrases, Vas a variable
over verbs, VP as a variable over verb phrases, and so on. By parallel
arguments, the logical labels should be a set of basic logical categories,
perhaps from a fixed range which is part of basic human mental
equipment, which might be regarded as variables over conceptual
representations of different types. We will use the pro-forms of English to
represent them: thus someone is a variable over conceptual representations
of people, something over conceptual representations of things, do
something over conceptual representations of actions, and so on.
To give the simplest possible illustration, sentence (56) has the
underlying tree structure (57), and its logical form has the structure (58):
(56) John invited Lucy
(57)
Sentence
VP
NP
John
NP
invited
Lucy
206

something is the case
(58)
someone
John
did something
invited
someone
Lucy
A propositional representation of the fact that John invited Lucy would
then carry, via the labels on the nodes of its tree structure, the information
that someone invited someone, that someone invited Lucy, that John
invited someone, that John did something, and so on.
On this approach, there is a clear sense in which the logical category
labels correspond to, and are indeed semantic interpretations of, syntactic
category labels of natural language (though there need not be a one-one
correspondence). As a result, a hearer who has made the anticipatory
syntactic hypothesis that, say, the words 'John invited' will be followed
by an NP, can by semantically interpreting this anticipatory syntactic
hypothesis derive the anticipatory logical hypothesis that John invited
someone. We see such hypotheses as playing a crucial role in disambiguation and reference assignment.
Let us assume that when he hears the word 'Jennifer' in (59), the hearer
accesses a range of possible referents for 'Jennifer' - that is, a set of
conceptual addresses with the word 'Jennifer' as part of their lexical entry
- and gains access in turn to a range of associated encylopaedic entries:
(59) Jennifer admitted STEALING.
On assigning 'Jennifer' to the syntactic category NP, he makes the
anticipatory syntactic hypothesis that it will be followed by a VP, which
yields by variable-substitution the anticipatory logical hypothesis (60):
(60) Jennifer did something.
Let us assume that he knows a Jennifer Smith and a Jennifer O'Hara. Our
hypothesis is that he now proceeds to make a tentative assignment of
reference to the expression 'Jennifer' by considering whether the information that Jennifer Smith did something or the information that Jennifer
O'Hara did something might be relevant to him in some context he
currently has accessible.
207
Suppose, for example, that he has accessible a context in which the

information that Jennifer O'Hara did something would be manifestly
relevant to him. Then by an argument that should by now be familiar, he
should assume that when the speaker said 'Jennifer' she was intending to
refer to J ennifer O'Hara. Otherwise she should have rephrased her
utterance to eliminate this interpretation. Moreover, he should assume
that the context in which he finds relevant the information that Jennifer
O'Hara did something will play a further role in the comprehension
process: otherwise the effort of accessing it will be wasted.
At this early stage there may be no obvious assignment of reference
which would make (60) relevant at all. Nonetheless, on some assignment
of reference it may raise a relevant question in the hearer's mind (where a
relevant question is a question the answer to which is certain or likely to be
relevant). A statement often raises a relevant question. For example, if I
tell you that I am unhappy, I will almost certainly make you wonder why.
By the same token, there may be some assignment of reference on which
(60), while not relevant itself, might raise a question such as (61 a) or (61 b)
which is relevant in some context currently accessible to the hearer:
(61) (a) What did Jennifer Smith do?
(b) What did Jennifer O'Hara do?
If so, then by a now familiar argument he should assume that this is a

question the speaker intended to raise, that the rest of the utterance will
answer it, and that the answer will be relevant in the context he has just
been encouraged to access.
Let us assume, then, that a reference has been tentatively assigned to the
expression 'Jennifer'. The next word to be processed is the verb
'admitted': this has two possible senses, confess to and let in, on both of
which it is transitive. The hearer can thus make the anticipatory syntactic
hypothesis that it will be followed by an NP and obtain, by variablesubstitution, the anticipatory logical hypotheses (62a-b):
(62) (a) Jennifer let someone in.
(b) J ennifer confessed to something.
The hearer, who now has access to the encyclopaedic entries for let in and
confess to, can proceed to disambiguate the verb 'admit' by asking himself
whether one of (62a-b), or one of the related questions (63a-b), is
manifestly relevant in some context which he currently has accessible:
(63) (a) Who did J ennifer let in?
(b) What did J ennifer confess to?
If so, then he should provisionally accept that interpretation and retain the
context for further processing.
208
Finally, if the speaker has achieved optimal relevance, the word

'stealing' should fall into a place already prepared for it during the
interpretation process. That is, it should answer any questions that have
been raised but left unanswered, and the answers should be relevant in a
context that the interpretation process has helped to prepare. For example,
if the hearer has accessed a context in which the question 'What did
Jennifer confess to?' would be relevant to him on some assignment of
reference to the expression 'J ennifer', then the utterance as a whole should
be interpretable as an answer to that question on that assignment of
reference.
If the speaker has done her job properly, the end of the utterance should
confirm all the provisional choices of content and context that have been
made along the route. On the other hand, if the end of the utterance does
not confirm these provisional choices, then identification of the speaker's
informative intention will involve an extra layer of inference in an
otherwise unchanged inference pattern. When communication is unproblematic, the hearer just takes for granted that the speaker has an adequate
appreciation of what would be relevant to him; when problems arise, the
hearer should try to find out under what mistaken image of him the
speaker could have thought that her utterance would be optimally
relevant.
Notice now that the correct anticipatory hypotheses, the ones that will
eventually be confirmed, are logically related to one another. In any pair
of such hypotheses, one is necessarily implied by the other. More
precisely, the set of anticipatory hypotheses forms a scale, in which each
member analytically implies the immediately preceding member and is
analytically implied by the immediately succeeding member. For (59), on
the interpretation we have just been discussing, the appropriate scale has
the three members (64a-c):
(64)(a) Jennifer did something/
What did J ennifer do?
(b) J ennifer confessed to something/
What did J ennifer confess to?
(c) J ennifer confessed to stealing.
Of these, the most general, (64a), will be recovered first, and as we have
seen, if the speaker has achieved optimal relevance, it should either be
relevant in its own right or raise a relevant question; the next to be
recovered will be (64b), which should either give a relevant answer to a
question raised by (64a) or raise a relevant question itself. Finally, (64c)
should give a relevant answer to the question raised by (64b), and might
itself raise the further question 'Why did Jennifer confess to stealing?',
thus preparing the ground for a subsequent utterance. The scale in (64)
thus acts as a skeleton around which the whole interpretation is built.
209 .
The scale in (64) contains a subset of the analytic implications of (59).

This subset is strictly ordered by the relation of analytic implication, as
described above. Moreover, given that focal stress in (59) falls on the last
word, 'stealing', this subset is related to the set of possible foci of the
utterance in an obvious way: take the propositional form of the utterance,
replace the focus by its logical label, and you get an implication in the
scale; all the implications in the scale, except the propositional form itself,
can be obtained in this way. Let us call such a strictly ordered subset of
analytic implications, determined by the placement of focal stress, a focal
scale. When focal stress falls on the last word of an utterance, as it does in
(59), the set of anticipatory logical hypotheses made during the interpretation process coincides with the focal scale.
As we have shown, implications in the focal scale of an utterance are not
all processed at once. The processing of each implication can contribute to
the overall relevance of the utterance in two ways: either by reducing the
effort needed to process it, or by increasing its contextual effects. Even if it
has no contextual effect in its own right, an implication can contribute to
relevance by giving direct access to a context in which effects can be
achieved, and thus reducing the processing effort needed to achieve these
effects. As for the contextual effects of an utterance, they may be obtained
in several steps, via different implications in the focal scale.
Let us say that when an implication in the focal scale of an utterance has
contextual effects of its own, and hence is relevant in its own right, it is a
foreground implication, and that otherwise it is a background implication.
Then the focus of an utterance will be the smallest syntactic constituent
whose replacement by a variable yields a background rather than a
foreground implication. For example, in (59) above the focus could be the
NP 'stealing', the VP 'admitted stealing' or the sentence as a whole. If (64a)
is relevant in its own right, it will be a foreground implication and the
focus will be the sentence as a whole; if (64a) is not relevant in its own right
but (64b) is, then the focus will be the VP 'admitted stealing'; and if (64b) is
not relevant in its own right, then the focus will be the NP 'stealing' itself.
An implication may contribute to relevance in both of the ways
described above: by giving access to a context in which further implications will have contextual effects, and by yielding contextual effects in its
own right. It is thus quite possible for the speaker not to know or care
exactly where the break between foreground and background will come.
In processing (59) above, for example, it might well be that if the hearer
accessed enough encyclopaedic information about Jennifer's character
and inclinations, he might find it relevant that she has confessed to
something. But maybe he is not prepared to put in that much effort. This
should make little difference to the speaker. As long as she has reason to
believe that at least one member of the focal scale will be relevant in its own
right, and in a context to which other members give access, she need not
210
care too much about where the cut-off point between foreground and
background will come. That is, she need have no specific intention about
which of the implications of her utterance are foreground and which
background (which are given and which new), contrary to what is
normally accepted in the literature.
We can also shed some light on the intuition that there is a gradient of
given and new information. Wherever the cut-off point between foreground and background comes, there is a clear sense in which (64b), for
example, simultaneously acts as a foreground implication in relation to
(64a), giving a partial answer to the question it:raises, and as a background
implication in relation to (64c), raising a question to which (64c) gives at
least a partial answer. As we have seen, even (64c), which is necessarily a
foreground implication, may simultaneously raise a background question
which some subsequent utterance (or a continuation of the same
utterance) will answer. Our distinction between foreground and background, like our notion of focus itself, is thus a purely functional one, and
should play no role in the linguistic description of sentences.
Different stress assignments induce different focal scales If the focally
stressed constituent were 'Jennifer', the focal scale for (59) would be (65):
(65) (a) Someone confessed to stealing/
Who confessed to stealing?
(b) J ennifer confessed to stealing.
If the focally stressed constituent were the verb 'admitted', the focal scale
would be (66):
(66) (a) Jennifer did something/
What did J ennifer do?
(b) J ennifer did something regarding stealing/
What did Jennifer do regarding stealing?
(c) J ennifer confessed to stealing.
(65) and (66) have the same logical properties as (64): each consists of a
series of logically related members, each member analytically implying the
immediately preceding member and being analytically implied by the
immediately following member. Moreover, each is obtainable by the same
general procedure: take the full propositional form of the utterance and
replace by a logical variable, first the interpretation of the focally stressed
constituent, then the interpretation of the next smallest syntactic constituent which contains the focally stressed constituent, and so on until there
are no more inclusive constituents to be replaced.
However, there is an important difference between (64) on the one hand
and (65) and (66) on the other, linked to the fact that, in the two latter
cases, the focal stress is not on the last word of the sentence. As a result, the
211
focal scale, which is determined by stress placement, cannot correspond,

at least not entirely, to the scale of anticipatory hypotheses, which is
determined by word order. It is hard to see how a speaker could lead her
hearer step by step, anticipatory hypothesis by anticipatory hypothesis,
up the scales in (65) and (66) in the way illustrated above for (64). For
example, unless the hearer already knew how the utterance was going to
end, he would be unable to make the anticipatory hypothesis (65a) on
presentation of the word 'J ennifer', or the anticipatory hypothesis (66b)
on presentation of the word 'admitted'. Though (64)-(66) have similar
logical properties and a similar relation to the syntactic structure of (59), in
normal circumstances only (64) is recoverable by a step-by-step series of
anticipatory hypotheses as the utterance proceeds.
The sense in which it is natural for focal stress to fall at the end of the
utterance, and hence for the background to be recovered before the
foreground, is the sense in which it is natural to raise a question before
answering it or to communicate a complex piece of information step by
step. However, departures from this pattern do occur, and these, too, may
be consistent with the principle of relevance. For example, when a
question has been explicitly or implicitly raised by the immediately
preceding utterance, it would be a waste of processing effort to repeat it.
Responses to explicit questions can thus be very fragmentary, and in a
non-fragmentary response the focus may precede the background, as
illustrated by the possible responses (67b-d) to the question in (67a):
(67) (a)
(b)
(c)
(d)
Peter:
Mary:
Mary:
Mary:
Who is the greatest English writer?

SHAKESPEARE is the greatest English writer.
SHAKESPEARE is.
SHAKESPEARE.
Each of these responses has the NP 'Shakespeare' as focus and the

assumption Someone is the greatest English writer as background. Only in
(67b) is this assumption made fully explicit, and here it follows the focus,
acting merely as a reminder, a confirmation of an interpretation which the
hearer should have been able to arrive at unaided.
There has been some discussion about whether the contrast between
examples such as (68b) and (69b) can be dealt with in purely pragmatic
terms: 1S
(68) (a) I'm sorry I'm late. (b) My CAR broke down.
(69) (a) I'm sorry I'm late. (b) My car was BOOBY-trapped.
The issue is whether in these examples stress has lost its natural
highlighting function and become subject to arbitrary linguistic constraints, or whether its position can be seen as following in some
212

interesting way from pragmatic principles. If a pragmatic account is
possible, we suggest that it might go as follows.
There is a quite deep-rooted intuition that in initially stressed examples
such as (68b), the material following the subject noun phrase is predictable
in some sense. The appropriate sense, we believe, is derivable from our
notions of weak implicature and weak communication. When the speaker
of (68a) apologises for being late, the hearer will naturally expect an
explanation. At the mention of the words 'my car' he will automatically
make the anticipatory hypothesis that something happened to her car
which caused her to be late, and access:the appropriate part of his
encyclopaedic entry for car. There he will find the highly accessible and
strongly confirmed assumptions that cars can break down, run out of
petrol and fail to start, all of which would reasonably account for the
speaker's failure to arrive on time.
Notice that the speaker could weakly implicate this range of excuses
without ever expressing it explicitly. Thus if she said, 'I'm sorry I'm late.
My damned car!' she would be understood precisely to have invoked one
of these predictable excuses. Thus, what follows the initial noun phrase in
these examples is already weakly communicated by the mention of the
initial noun phrase alone, and the contextual effects on which the main
relevance of the utterance depends can be calculated on the basis of the
resulting anticipatory hypothesis.
Attempts at explaining the contrast between examples like (68b) and
(69b) have rarely gone beyond vague appeals to 'relative semantic weight'
or 'relative newsworthiness'. In our framework it is possible to do rather
better. To the hearer of (69b), even after he has made the anticipatory
hypothesis that the speaker is late because something happened to her car,
the information that the car was booby-trapped would still be quite
relevant. In other words, the material following the initially unstressed
noun phrase in (69b) has significant contextual effects in a context created
by the processing of the initial noun phrase, whereas the material
following the initially stressed noun phrase in (68b) does not. As a result,
the material following the initial noun phrase in (69b) should be focally
stressed, whereas the material following the initial noun phrase in (68b)
should not. Along these lines, the contrast between examples such as (68)
and (69) might be approached without appealing to any ad hoc stress
assignment rules.
At least in the recent literature on generative grammar, contrastive
stress has generally been treated as a non-linguistic or paralinguistic
phenomenon, not subject to special phonological constraints. This fits
well with the view of stress as a purely natural device for pinpointing some
noteworthy aspect of an utterance. At first sight, the fact that contrastive
stress works differently in different languages presents problems for this
213
approach. It is used much more freely in English than in French: for

example, whereas (70) sounds entirely natural, its French counterpart
(71 b) could only occur naturally as echoing a preceding utterance such as
(71a):
(70) YOU must do the washing up.
(71) (a) He: 11 faut que vous fassiez la vaisselle.
(b) She: Non, il faut que VOUS fassiez la vaisselle.
)
)
s
e
c
cs
le
re
is
However, this is not a particularly compelling objection unless it can be

shown that variations in contrastive stress can not be accounted for in
processing terms. If the same stress pattern may have a higher processing
cost in one language than another, or when realised by one expression
rather than another, then a speaker aiming at optimal relevance should use
the costlier stress pattern more sparingly.
It seems clear that in a language like French, with its relatively flat
intonation contour and strongly preferred final placement of focal stress,
the non-final use of contrastive stress will cause a much greater disruption
to the intonation contour than in a language like English, with its
relatively variable intonation contour and freer placement of focal stress.
Greater disruption implies greater processing effort and, other things
being equal, lowered acceptability. We would predict, then, that
utterance-internal contrastive stress in a language like French would be
acceptable only in an echoic exchange such as (71) above, where
parallelisms in linguistic form and pragmatic interpretation reduce overall
processing effort, and where it offers a particularly economical way of
pinpointing an intended range of contextual effects.
We are suggesting, then, that stress placement, like other stylistic
features, should be looked at in terms of processing effort. The fact that
contrastive stress is a natural highlighting device need not prevent it from
being more costly to use in some circumstances than in others, just as
pointing, another natural highlighting device, may have greater social
costs attached to it in some circumstances than in others. 16 This suggests,
then, an interesting approach to cross-linguistic variation in stress
patterns, and one that we feel might shed light on a range of noncontrastive data which have been the subject of much interesting recent
research. 17
Returning now to our original hypothesis of a natural linkage between
linguistic structure and pragmatic effects, let us show how examples
(52)-(54) above would be dealt with in the framework sketched in this
section. Consider (52a-b):
(52) (a) Bill's twin sister lives in BERLIN.
(b) Bill has a twin sister who lives in BERLIN.
214

These examples are standardly described in terms of a presuppositionassertion distinction: (52a) presupposes what (52b) asserts - namely, that
Bill has a twin sister. Two types of intuition are called on to justify the
distinction. Suppose that Bill has no twin sister and that the hearer knows
it. Then there is one intuition we agree with, which indeed seems hardly
controvertible: that (52a) would be perceived as much more seriously
defective than (52b). It has been claimed that there is a stronger intuition:
that if Bill has no twin sister, then regardless of the hearer's belief about the
matter, (52a) does not express a proposition at all. This second intuition
we dispute, and since it is in any case marginal to pragmatics we do not
propose to discuss it here. 18 The first intuition, however, follows directly
from our framework.
Let us assume that both (52a) and (52b) analytically imply (72):
(72) Bill has a twin sister.
Then the two utterances share their truth conditions. However, in (52b),
(72) is in the focal scale, which in this case corresponds to the scale of
anticipatory hypotheses. It acts as a development of the preceding
hypothesis, 'Bill has something', or, equivalently, as an answer to the
question, 'What does Bill have?'; and it is at least part of a relevant answer
to this presumably relevant question. If the only contribution of the
information that Bill has a twin sister were to make immediately accessible
an existing conceptual address for this twin sister, then the more
economical (52a) should be preferred. In (52a), (72) is not an implication in
the focal scale, and, correlatively, it does not answer any suggested
question. In fact, the first relevant question suggested by the focal scale of
(52a) is 'What does Bill's twin sister do?' When should a speaker aiming at
optimal relevance prefer (52b) to (52a)? When the information that Bill has
a twin sister is relevant enough in its own right. Hence a speaker who
regards the assumption that Bill has a twin sister as neither manifest, nor
manifestly plausible to the hearer, should choose (52b), and a hearer who
rejects this assumption would regard (52a) as much more seriously
defective than (52b).
As Straws on (1964b) has noted, there is also a subtler range of intuitions
having to do with the presuppositional effects of referential expressions in
different syntactic positions. Thus, compare (73a) and (73b):
(73) (a) The King of France visited the EXHIBITION.
(b) The exhibition was visited by the King of FRANCE.
According to Strawson, if there was an identifiable exhibition but no
identifiable King of France, (73b) would succeed in making an assertion,
though a mistaken one, whereas (73a) would make no assertion at all. In
our framework, these intuitions are predictable in terms of the distinction
215
between foreground and background implications. The strongest presuppositional effects are carried by analytic implications of background
implications. Thus, if the focus in (73a) is 'the Exhibition', or 'visited the
Exhibition', and the background is The King of France did something, or
The King of France visited something, the information that there is a King
of France will be analytically implied by the background, and a hearer
who rejects it will be unable to access a context in which the utterance
would be relevant at all. By contrast, if the focus in (73 b) is 'the King of
France', or 'was visited by the King of France', and the background is The
Exhibition had some property, or The Exhibition was visited by someone,
then at least the hearer will be able to access the appropriate context and
see what sort of contextual effects the speaker must have had in mind.
Hence the intuition that in this case the consequences of reference failure
are less dramatic.
Turning to examples (53a-c), triples of this type are often seen as
showing the need for two separate distinctions, one based on left-right
word order and the other on intonational prominence:
(53)(a) It rained on MONDAY.
(b) On Monday it RAINED.
(c) On MONDAY it rained.
Thus Halliday (1967-8) distinguishes thematic or textual structure, based
on left-right word order, from informational structure, based on intonational prominence. He defines the theme as the leftmost syntactic
constituent in the sentence and the rheme as everything that follows. The
theme-rheme distinction, like the focus-presupposition distinction, is
often seen as genuinely linguistic: thus Brown and Yule (1983: 133) claim
that 'theme is a formal category in the analysis of sentences'. In our
framework the differences between (53a), (53 b) and (53c) can be
accounted for without introducing theme as a formal category at all.
We have seen that (53a) has a range of possible foci: 'Monday', 'on
Monday' and the sentence as a whole. It is thus construable as an answer to
the questions, 'On what day did it rain?', 'When did it rain?' and 'What
happened?' The effect of (53 b) and (53c) is to modify this range of possible
interpretations. By the time the hearer of (53 b) has processed the words
'On Monday', he knows that there is some question about what happened
on Monday which the speaker thinks is relevant to him. In other words,
the effect of fronting the unstressed constituent 'On Monday' is to force it
into the background. By the same token, by the time the hearer of (53c)
has processed the words 'On Monday', he should know that they give the
answer to some question which he should at this point be able to access for
himself. In other words, the effect of fronting the stressed constituent 'On
Monday' is to select it as focus. Sentences such as (53b) and (53c) may
216
involve slightly greater processing costs than (53a); if so, this would be the
price paid for fixing an adverbial expression in the background while
retaining utterance-final stress, or of pinpointing the focus more precisely
than its normal syntactic position would permit. However, the special
effects of such structures arise simply from the interaction between
syntax, stress assignment and the principle of relevance. While the
theme-rheme distinction may be a valuable way of highlighting intuitions,
it has no place in the technical descriptive vocabulary of either linguistics
or pragmatics.
The classic paper on the topic-comment distinction is Reinhart (1981).
She defines the 'sentence topic' as a syntactic constituent, explicitly
present in the sentence, whose referent the sentence is about; many
authors also appeal to a vaguer notion of 'discourse topic'. In general,
sentence topics will be both unstressed and early in the word order. Thus
in (54a) the sentence topic is John-Paul the Second, in (54b) it is the present
Pope:
(54)(a) John-Paul the Second is the present POPE.
(b) The present Pope is John-Paul the SECOND.
(c) It is John-Paul the SECOND who is the present Pope.
As regards the pragmatic role of topics, there is a general agreement that
their function is to provide access to what in our terms would be
contextual information crucial to the comprehension process. Thus the
classic discourse topics are titles and picture captions, whose role is
precisely to give access to encyclopaedic information crucial to the
comprehension of the accompanying texts or pictures; by the same token,
sentence topics are generally unstressed syntactic constituents occurring
early in the utterance, whose function in our framework is to give access to
encyclopaedic information which the speaker regards as crucial to the
interpretation process.
One reason for looking seriously at the literature on topics is that it is
often claimed that the most basic notion of relevance, the one it is most
important to define, is that of relevance to a topic. Thus Brown and Yule
(1983: 68) comment that though the notion of topic is 'very difficult to pin
down' it is nonetheless 'essential to concepts such as relevance and
coherence'. Given the role of topics in providing access to contexts, these
comments are not too surprising. To the extent that an utterance is
relevant (in our sense) in a homogeneous context derivable from a single
encyclopaedic entry, it will be topic relevant (in a derivative sense), the
topic being simply the conceptual address associated with that encyclopaedic entry. However, in our framework an utterance may also be
relevant in a non-homogeneous context - that is, a context derived from a
variety of encyclopaedic and environmental sources - in which it should
Implicatures and style
217
be hard to elicit systematic judgements about its topic relevance.

Topic-relevant utterances are only a subset of relevant utterances, and it is
the notion of topic relevance which is derivative. 19
As regards the given-new and focus-presupposition distinctions, we
have little to add to what has already been said. In our framework,
background information is information that contributes only indirectly to
relevance, by reducing the processing effort required; it need be neither
given nor presupposed. Foreground information is information that is
relevant in its own right by having contextual effects; it need not be new. 20
However, the fundamental difference between our foregroundbackground distinction and the given-new and focus-presupposition
distinctions is in their theoretical status. The given-new and focuspresupposition distinctions are typically regarded as part of the basic
machinery of linguistic and/or pragmatic theory. Our foregroundbackground distinction, by contrast, has no role at all to play in linguistic
theory, and in pragmatics it is simply a descriptive label used to distinguish
twO complementary and independently necessary aspects of the interpretation process. We do not assume, that is, that competent speakers have
to have, either built into their grammar or built into their inferential
abilities, any notion of background and foreground. Backgrounding and
foregrounding arise as automatic effects of the hearer's tendency to
maximise relevance, and of the speaker's exploitation of that tendency.
The main argument of this section has been as follows. Given that
utterances have constituent structure, internal order and focal stress, and
given that they are processed over time, the most cost-efficient way of
exploiting these structural features will give rise to a variety of pragmatic
effects. There is a natural linkage between linguistic structure and
pragmatic interpretation, and no need for any special pragmatic conventions or interpretation rules: the speaker merely adapts her utterance to
the way the hearer is going to process it anyhow, given the existing
structural and temporal constraints. 21
6 Implicatures and style: poetic effects

It is sometimes said that style is the man. We would rather say that style is
the relationship. From the style of a communication it is possible to infer
such things as what the speaker takes to be the hearer's cognitive capacities
and level of attention, how much help or guidance she is prepared to give
him in processing her utterance, the degree of complicity between them,
their emotional closeness or distance. In other words, a speaker not only
aims to enlarge the mutual cognitive environment she shares with the
218
hearer; she also assumes a certain degree of mutuality, which is indicated,

and sometimes communicated, by her style.
Choice of style is something that no speaker or writer can avoid. In
aiming at relevance, the speaker must make some assumptions about the
hearer's cognitive abilities and contextual resources, which will necessarily be reflected in the way she communicates, and in particular in what she
chooses to make explicit and what she chooses to leave implicit. Compare
(74a-c):
(74) (a) Only amateurs can compete in the Olympics.
(b) The Olympic Games is an international sporting competition
held every four years. Only amateurs can compete.
(c) The Olympic Games is an international sporting competition held
every four years. Only amateurs - that is, people who receive no
payment for their sporting activities - can compete in the
Olympic Games. Professionals - that is, people who receive some
payment for their sporting activities - are not allowed to compete
in the Olympic Games.
These utterances differ not so much in their import as in the amount of
help they give the hearer in recovering it. What the speaker of (74a) trusts
the hearer to know about the Olympics is stated explicitly in (74b) and
(74c). What the speaker of (74a) and (74b) trusts the hearer to know about
amateur status is stipulated explicitly in (74c). The style of (74c) is heavier
than that of (74b), which is itself heavier than that of (74a), and this is due
to the difference in reliance on the hearer's ability to recover implicit
Import.
A speaker aiming at optimal relevance will leave implicit everything her
hearer can be trusted to supply with less effort than would be needed to
process an explicit prompt. The more information she leaves implicit, the
greater the degree of mutual understanding she makes it manifest that she
takes to exist between her and her hearer. Of course, if she overestimates
this degree of mutual understanding, there is a risk of making her
utterance harder or even impossible to understand. It is not always easy to
strike the correct balance: even a slight mismatch between speaker's
estimate and hearer's abilities may make what was merely intended to be
helpful seem patronising or positively offensive to the hearer. What is
important, however, is that the speaker must choose some form in which
to convey her intended message, and that the form she chooses cannot but
reveal her assumptions about the hearer's contextual resources and
processing abilities. There is no entirely neutral style.
Another dimension along which styles may vary is in the degree to
which they constrain or guide the hearer's search for relevance. Compare
(75b-d) as answers to question (75a):
219
(75) (a) Peter: Is Jack a good sailor?

(b) Mary: Yes, he is.
(c) Mary: ALL the English are good sailors.
(d) Mary: He's English.
As we saw in section 4, whereas a direct answer leaves the hearer free to
process the information offered in whatever way he likes, an indirect
answer suggests a particular line of processing in the computation of
contextual effects. In saying (75c), for example, Mary not only expects
Peter to access and use the assumption that Jack is English and infer that
Jack is a good sailor; she also encourages him to speculate on, to derive
some additional conclusions from, the assumption that the English are
good sailors. In saying (75d), by contrast, she behaves as if the assumption
that all the English are good sailors were mutually manifest to her and
Peter, and more manifest than the assumption that Jack is English. There
could be circumstances in which the main relevance of (75d) came not
from the strongly implicated conclusion that Jack is a good sailor, but
from the fact that Mary, by treating it as mutually manifest that all the
English are good sailors, has made mutually manifest her intention to
make manifest that she assumes she shares with Peter a sense of national
pride.
Style arises, we maintain, in the pursuit of relevance. The classical
figures of speech were defined in terms of formal features which mayor
may not have the expected stylistic effects. Consider epizeuxis or
repetition, for example. The effects of repetition on utterance interpretation are by no means constant. Compare the following:
(76)
(77)
(78)
(79)
(80)
(81)
Here's a red sock, here's a red sock, here's a blue sock.

We went for a long, long walk.
There were houses, houses everywhere.
I shall never, never smoke again.
There's a fox, a fox in the garden.
My childhood days are gone, gone.
In circumstances that are easy to imagine, (76) might convey that there are
two red socks; (77) that the speaker went for a very long walk; (78) that
there were a great many houses, (79) that the speaker will definitely never
smoke again, (80) that the speaker was excited about the fox in the garden;
and (81) that she was moved by the disappearance of her childhood days.
Thus the 'emphatic' effects of repetition are worked out in different ways
for different examples. In particular, they may be reflected in the
propositional content of the utterance, as in (76)-(78), in the speaker's
degree of commitment to that propositional content, as in (79), or in some
other expression of the speaker's attitude, as in (80) and (81).

One way of accounting for this variation would be to set up specific
principles of semantic or pragmatic intepretation, so that, for example, the
first of two repeated scalar adjectives would be interpreted as meaning
very, the first of two repeated plural nouns would be interpreted as
meaning many, and so on. However, in the case of (80), and especially (81),
it is hard to think of propositional paraphrases that would adequately
capture their import. These utterances as it were exhibit rather than
merely describe the speaker's mental or emotional state: they give rise to
non-propositional effects which would be lost under paraphrase. Thus the
idea of ad hoc case-by-case semantic or pragmatic treatment of repetition
seems to have little to recommend it.
Another possibility would be to show that the effects of repetition
follow from more general psychological principles, perhaps from some
universal set of cognitive strategies towards repetitive inputs in nature. On
the face of it, though, it is hard to see how two sheep, or a flock of sheep,
could be construed as a moving, exciting or emphatic version of a single
sheep. Moreover, examples (80) and (81) again present problems for such
an approach.
From the point of view of relevance theory, both of these approaches
are superfluous, anyhow, since the interpretations of (76)-(81) follow
automatically from the principle of relevance. Within our framework, the
task of the hearer faced with these utterances is to reconcile the fact that a
certain expression has been repeated with the assumption that optimal
relevance has been aimed at. Clearly, the extra linguistic processing effort
incurred by the repetition must be outweighed by some increase in
contextual effects triggered by the repetition itself. The different interpretations of (76)-(81) simply illustrate the different ways in which such
an increase can be achieved.
With (76) it would be consistent with the principle of relevance to
assume that the two occurrences of 'Here's a red sock' refer to numerically
distinct objects: hence (76) is naturally understood as meaning that there
are two red socks. With (77) it would be consistent with the principle of
relevance to assume that the speaker wanted to indicate that the walk was
longer than the hearer would otherwise have thought: in other words, that
it was a very long walk. With (78) it would be consistent with the principle
of relevance to assume that the speaker wanted to indicate that there were
more houses than the hearer would otherwise have thought: in other
words, that there were a great many houses. In each of these cases the
repetition modifies the propositional form and hence the explicatures of
the utterance, and achieves extra contextual effects thereby.
None of these lines of interpretation is available for (79). Here It would
be consistent with the principle of relevance to assume that the speaker
attaches a higher confirmation value to the assumption expressed than the
220
221
hearer would otherwise have thought. Realising that her utterance will be
sceptically received, she repeats the word 'never', the likely target of the
scepticism, to convince the hearer that she means what she says. In other
words, 'never, never' is here similar in import to 'definitely never', and
reflects the speaker's degree of commitment to the assumption expressed.
This strengthens the explicature and all its contextual implications,
thereby increasing the contextual effects of the utterance.
With (80) and (81), none of the above interpretations works well. No
increase in effect is likely to be achieved either by enriching the
propositional form, or by strengthening the implicatures. We want to
suggest that in these cases, the repetition should yield an increase in
contextual effects by encouraging the hearer to extend the context and
thereby add further implicatures. The repetition in (80) cannot be
accounted for by assuming that there are several foxes in the garden, or by
strengthening the assumption that there is a fox. Instead, the hearer of (80)
is being encouraged to dig deeper into his encyclopaedic entry for fox,
with a guarantee that the extra processing effort will be outweighed by a
gain in contextual effects: the fact that there's a fox in the garden is
presented as more relevant than the hearer would have spontaneously
realised.
Similarly, the repetition in (81) cannot be accounted for by assuming
that the speaker's childhood days are longer gone, or more definitely
gone, than might otherwise have been assumed, so if the presumption of
relevance is to be confirmed, then the repetition of 'gone' must be
interpreted as an encouragement to expand the context. There is a
difference between (80) and (81), though. Paying attention to the fact that
there is a fox in the garden, and making the effort to remember basic facts
about foxes, is likely to yield some strong and predictable contextual
implications, such as 'The chickens are in danger'. These strong implications are likely to be interpreted as strong implicatures of the utterance. In
the case of (81), the extra relevance is more likely to be achieved by a more
diversified expansion of the context and by a wider array of weaker
implicatures. In other words, the hearer is encouraged to be imaginative
and to take a large share of responsibility in imagining what it may be for
the speaker to be way past her youth.
Compare the interpretation of (81) and (82):
(81) My childhood days are gone, gone.
(82) My childhood days are gone.
Our suggestion is not that, in a given context, (81) has contextual

implications that (82) lacks: the hearer of either utterance is free to derive
as many consequences as he pleases from the fact that the speaker's
childhood days are gone. What (81) has is more implicatures than (82): that
222
is, more contextual assumptions and implications which receive some

degree of backing from the speaker. To justify the repetition of 'gone', the
hearer must think of all the implicatures that the speaker could reasonably
have expected him to derive from (82), and then assume that there is a
whole range of still further premises and conclusions which the speaker
wants to back. For this, the hearer must expand the context. As a result
(81) might suggest, say, that the speaker is experiencing a torrent of
memories which the hearer is being trusted to imagine for himself. What
look like non-propositional effects associated with the expression of
attitudes, feelings and states of mind can \:se approached in terms of the
notion of weak implicature developed in section 4.
Let us give the name poetic effect to the peculiar effect of an utterance
which achieves most of its relevance through a wide array of weak
implicatures. Generally, the most striking examples of a particular figure,
the ones singled out for attention by rhetoricians and students of style, are
those which have poetic effects in this sense. These poetic effects are then
attributed to the syntactic or phonological construction in question.
However, as the above examples show, a repetitive syntactic pattern does
not invariably give rise to noticeable stylistic effects. The same is true of all
the figures of style identified by classical rhetoric.
Consider, for example, the syntactic construction that modern syntacticians call gapping and classical rhetoricians call zeugma, as illustrated in
(83)-(85):
(83) Mary went on holiday to the mountains, Joan to the sea, and Lily to
the country.
(84) Mary lives in Oxford, J oan in York, and Lily in a skyscraper.
(85) Mary came with Peter, Joan with Bob, and Lily with a sad smile on
her face.
In each of these examples there are clear syntactic, semantic and
phonological parallelisms. These reinforce the hearer's natural tendency
to reduce processing effort by looking for matching parallelisms in
propositional form and implicatures. In (83), for instance, the missing VP
in the second and third clause can safely be assumed to be 'went on
holiday'. Moreover, the same easily accessible context - scenarios of
typical holidays - enables the three clauses to yield parallel contextual
effects, with some conclusions true of Mary, Joan and Lily, and others
contrasting their respective holidays on fairly standard dimensions of
comparison. A speaker aiming at optimal relevance would deliberately
introduce such linguistic parallelisms only if she expected them to lead to a
reduction in the hearer's processing effort, and in particular, if she thought
that the search for parallel contexts and contextual effects would be
223
rewarded. Otherwise, the parallelisms might misdirect the hearer's effort,

thus increasing instead of reducing it. Thus, to the extent that it reflects the
speaker's deliberate choice, the form of (83)-(85) indicates to the hearer
that the search for parallel contexts and contextual implications will be
successful.
The parallelism in (83) achieves no striking stylistic effect. In the case of
(84), and even more of (85), it does. Here is an explanation: in the case of
(83), the syntactic parallelism is matched by semantic parallelism, and
parallel contextual effects are easily achieved in a largely common context.
The parallelism of (83), therefore, contributes to relevance merely by
reducing processing effort, and not by creating special contextual effects.
In the case of (84) and (85), the syntactic parallelism is not matched by a
similar semantic parallelism in the third clause: 'a skyscraper' does not
belong with 'Oxford' and 'York'; 'a sad smile on her face' does not belong
with 'Peter' and 'Bob'. The syntactic parallelism, however, is too salient to
be accidental or to go unnoticed; it is strong enough to trigger parallel
processing in spite of the partial semantic divergence. The problem is then
one of finding a context in which all three clauses have parallel contextual
effects. This requires an effort of imagination: the hearer has to bring
together relatively unrelated encyclopaedic entries and construct nonstereotypical assumptions.
In the case of (84), the hearer's task is to find a set of assumptions in the
context of which the facts that Mary lives in Oxford, J oan lives in York
and Lily lives in a skycraper have either identical or directly contrasting
implications. Some basic facts about Oxford, York and skyscrapers
suggest the conclusion that Mary and Joan do not live in skyscrapers, and
that Lily does not live in an old town. Still, these conclusions could have
been derived more cheaply if the speaker had named the town where Lily
lived, or the type of building where Mary and Joan lived. If the overall
interpretation is to be consistent with the principle of relevance, the
speaker must be credited with implicating more than that: for example,
she might have been trying to convey a variety of weak implicatures
showing that the way Mary and Joan live is more affected by the kind of
town they live in than by the kind of building they live in, while the reverse
is true of Lily's way of life.
In the case of (85), the hearer's task is to find a set of assumptions in the
context of which the facts that Mary came with Peter, Joan with Bob, and
Lily with a sad smile on her face have either identical or directly
contrasting implications. What might be suggested is that Lily had no one
to come with, that she was sad because she had no one to come with, that
there was a whole story behind her sad smile in which Mary, Peter, Joan
and Bob were somehow involved, and which an imaginative hearer could
224
spell out along a whole variety of lines. In this way the required
parallelisms of context and contextual effects could be maintained. The
result would be a wide range of fairly weak implicatures.
In (83), (84) and (85), because of the form of the utterance, the search
for an interpretation consistent with the principle of relevance induces a
certain processing strategy; in the case of (83) this strategy yields an
unremarkable interpretation; the contribution to relevance made by the
form of the utterance is merely to reduce processing effort. In the case of
(84), and even more so of (85), this strategy takes the hearer beyond
standard contexts and premises, and yields typical poetic effects.
How do poetic effects affect the mutual cognitive environment of
speaker and hearer? They do not add entirely new assumptions which are
strongly manifest in this environment. Instead, they marginally increase
the manifestness of a great many weakly manifest assumptions. In other
words, poetic effects create common impressions rather than common
knowledge. Utterances with poetic effects can be used precisely to create
this sense of apparently affective rather than cognitive mutuality. What we
are suggesting is that, if you look at these affective effects through the
microscope of relevance theory, you see a wide array of minute cognitive
effects.
Poetic effects, we claim, result from the accessing of a large array of very
weak implicatures in the otherwise ordinary pursuit of relevance. Stylistic
differences are just differences in the way relevance is achieved. One way
in which styles may differ is in their greater or lesser reliance on poetic
effects, just as they may differ in their greater or lesser reliance on
implicature and in the way they exploit the backgrounding and foregrounding of information in their explicatures.
7 Descriptive and interpretive dimensions

of language use
So far, we have restricted our attention to ordinary assertions: utterances
whose main explicature is simply their propositional form. In many cases
- perhaps most - the propositional form of an utterance is not an
explicature at all. This is true of tropes on the one hand, and of nonassertive speech acts on the other. Normally, though, these two kinds of
utterance are not regarded as particularly closely related.
The tropes are traditionally analysed as involving the substitution of a
figurative for a literal meaning. Consider the ironical (86):
(86) (a) Peter is quite well-read. (b) He's even heard of Shakespeare.
The propositional form of (86a) is the assumption that Peter is quite
Descriptive and interpretive uses
225
well-read. However, this is not an assumption that the speaker wants to

make manifest; it is not an explicature. The only obvious explicature of
(86a) is (87), where as we have already seen, one can say that P without
asserting or explicating that P.
(87) The speaker is saying that Peter is quite well-read.
Similarly, the propositional form of the metaphorical (88)
explicature:
IS
not an
(88) This room is a pigsty.

The speaker does not expect her hearer to start looking around for pigs.
The only obvious explicature of (88) is (89):
(89) The speaker is saying that this room is a pigsty.
The problem with figurative utterances such as (86a) and (88), which seem
to have no other explicatures than reports such as (87) and (89), is to
explain how they could be relevant at all.
In the case of speech acts other than assertions, the propositional form
of the utterance is not an explicature either. Consider a yes-no question
such as (90):
(90) Is Jill coming to the party?
The propositional form of (90) is (91):
(91) Jill is coming to the party.
However, if (90) is a genuine question, the speaker's intention is not to
communicate thatJill is coming to the party, but to find out whether she is.
The propositional form (91) has to be integrated into an assumption
schema such as (92) to yield (93), the explicature of (90):
(92) The speaker is asking whether it is true that - - (93) The speaker is asking whether it is true that J ill is coming to the party.
Similarly, the propositional form of the request in (94) is (95):
(94) Close the door, please.
(95) [The hearer] will close the door immediately.
However, the speaker's intention is clearly not to communicate to the
hearer that he will close the door immediately. The propositional form
(95) has to be integrated into an assumption schema such as (96) to yield
(97) as an explicature of (94):
(96) The speaker is telling the hearer to make it true that - - .
(97) The speaker is telling the hearer to make it true that he will close the
door immediately.
226
There is a considerable literature on illocutionary forces and speech

acts, and an even more considerable one on tropes. In both cases the
central concern has been with problems of classification, and little is
offered in the way of explanation. Despite this superficial resemblance,
there is very little overlap between work on illocutionary forces and work
on tropes, as if it went without saying that they are essentially different
aspects of language use. We do not share this view of illocutionary forces
and tropes as defining two homogeneous and radically distinct domains.
We would like to suggest a different and more integrated approach, based
on a fundamental distinction between interpretation and description. 22
This distinction is not an ad hoc piece of extra machinery introduced to
account for tropes and illocutionary forces. It follows quite naturally from
the relevance-based account of ostensive-inferential communication
which we have been trying to develop. In this section, we will introduce
and illustrate this distinction. We will then use it to explain metaphor in
section 8 and irony in section 9, and to look at speech acts and
illocutionary forces in a new light in section 10.
Most stimuli used in ostensive communication are representations
(public rather than mental representations, of course). This is true not
only of linguistic utterances, but of many other kinds of ostensive stimuli
as well. Relevance theory provides a straightforward explanation of this
fact, and without having to appeal to any ad hoc rule, constraint, or
principle: for example, a tacit convention to the effect that to represent
some state of affairs is to suggest that it exists (to which there are countless
counterexamples anyhow).
Identifying a stimulus, an ostensive stimulus in particular, involves
entertaining a logical form, a structured set of concepts. As we have
seen, concepts give access to encyclopaedic entries, and a logical form can
be used as an assumption schema. Given the principle of relevance, and in
particular the presumption that on the effort side the stimulus is the best
the communicator could have chosen, the addressee of an act of
communication is entitled to assume that, to recover the intended
interpretation, he must use the assumption schema suggested by the
logical form he is entertaining, and the encyclopaedic entries made
accessible by its constituent concepts.
A recognisable representation can be used to draw the audience's
attention to concepts and assumption schemas which are not instantiated
in the immediately perceptible environment. If you want someone to
think of a dog when there are no dogs around to point at, use a
representation of a dog: a drawing, a dog-like posture, an imitation of a
bark, the word 'dog', the word 'chien'. If you want someone to think of a
dog biting, use a representation, verbal or visual, of a dog biting. Since this
is ostensive behaviour, your addressee will assume that you are communi-
227
eating, that the information you are communicating is worth his attention,
that the stimulus you use is economical, that you are therefore not
gratuitously making him entertain the mental representation of a dog
biting, and that the first inferable interpretation consistent with these
assumptions should be the right one.
In appropriate conditions, any natural or artificial phenomenon in the
world can be used as a representation of some other phenomenon which it
resembles in some respects. Having climbed the walls of the villa, the first
thief silently imitates a dog biting to warn her accomplice at the foot of the
wall. You ask me what is the shape of Brazil, and for answer I point to an
appropriately shaped cloud in the sky. Mary wants to communicate to
Peter that she would like to leave the party, and she mimics the act of
driving.
Utterances can be used as representations in another way, too: not in
virtue of resembling some phenomenon, but in virtue of having a
propositional form which is true of some actual or conceivable state of
affairs. In the case of an assertion, for instance, the propositional form of
the utterance is used to represent some state of affairs in the real world; in
the case of a request, the propositional form of the utterance is used to
represent a desirable state of affairs. However, utterances are also
phenomena, and like all phenomena they can be used to represent
something they resemble. This possibility is often overlooked by
theorists, and even when it is not, we want to argue that the role it plays in
verbal communication is grossly underestimated.
Consider the following dialogue:
(98) Peter: What language did you speak to the inn-keeper?
Mary: Bonjour, comment allez-vous, bien, merci, et vous?
Mary communicates that she spoke French to the inn-keeper, not by

asserting it but by imitating the fact she wants to make manifest. Her
utterance is produced because it resembles the phenomenon of her
speaking French to the inn-keeper, even though its propositional form in
no way describes this phenomenon. However, as in example (6) of the
stammerer showing off, Mary's answer is hardly a case of true linguistic
commumcation.
Now consider dialogue (99):
(99) Peter: And what did the inn-keeper say?
Mary: Je l'ai cherche partout!
Here again, Mary is not communicating the propositional form of her

utterance. She is using this utterance because it resembles the inn-keeper's
utterance. It resembles that utterance because it is a token of the same
sentence: it is a direct quotation. A direct quotation has a linguistic
228
structure, including a semantic structure, and when it is used to make this

semantic structure manifest, it falls within the domain of verbal communication proper. Direct quotations are the most obvious examples of
utterances used to represent not what they describe but what they
resemble. They are by no means the only examples.
Consider another version of the dialogue between Peter and Mary:
Mary: I looked for it everywhere.
This time, Mary's utterance is a translation of the inn-keeper's utterance.

Again, it is used to represent what it resembles: it resembles the
inn-keeper's utterance because it has the same semantic structure.
Consider now,
Mary: He has looked for your wallet everywhere. I don't believe
him, though.
In (101), the first sentence of Mary's utterance is a representation of the

inn-keeper's utterance, though it is neither a direct quotation nor a
translation. How does Mary's utterance resemble the inn-keeper's? The
two utterances have different semantic structures, since the inn-keeper
used a first-person rather than a third-person pronoun to refer to himself,
and a third-person pronoun rather than a definite description to refer to
Peter's wallet. What the two utterances have in common is their
propositional form.
Now suppose that instead of the single sentence 'Je l'ai cherche
partoutl', the inn-keeper had produced a long speech which contained
neither this nor any closely similar sentence. Consider the following:
Mary: That he has looked for your wallet everywhere.
Once more, Mary's utterance is used to represent what it resembles,

namely the inn-keeper's speech. But what is the nature of the resemblance
this time? The linguistic structures are different, the semantic structures
are different, the propositional forms are different. However, if Mary's
summary is a faithful one, the propositional forms, though different, must
resemble one another: they must share some logical properties, have
partly identical contextual implications in some contexts, for instance.
Any representation with a propositional form, and in particular any
utterance, can be used to represent things in two ways. It can represent
some state of affairs in virtue of its propositional form being true of that
state of affairs; in this case we will say that the representation is a
description, or that it is used descriptively. Or it can represent some other
229
representation which also has a propositional form - a thought, for

instance - in virtue of a resemblance between the two propositional forms;
in this case we will say that the first representation is an interpretation of
the second one, or that it is used interpretively.
How closely must the propositional forms of two representations
resemble each other if the one is to be an interpretation of the other? We
will show that the answer to this question varies from case to case, but
invariably follows from the principle of relevance. What we want to point
out here is that while there may be a minimal degree of resemblance below
which no interpretive use is possible, there need not be a maximal degree
above which resemblance is replaced by identity and interpretation by
reproduction. Identity is a limiting case of resemblance; reproduction is a
limiting case of interpretation. When one representation is used to
represent another which has exactly the same propositional form, as in
example (101), this is merely a limiting case of interpretation.
The only generally acknowledged interpretive use of utterances is the
reporting of speech or thought: when one utterance is used to report
another utterance, as in examples (99)-(102), or a thought, as in (103):
(103) Mary: We won't bother go

safely keep the wallet.
to
the police, he thinks, and so he can
Here, Mary's utterance, with the exception of the parenthetical 'he

thinks', is used to report a thought which Mary attributes to the
inn-keeper.
There are other interpretive uses of utterances apart from reports of
speech or thought. Consider assumption (104):
(104) There is a prime number greater than 8,364,357 and smaller than
8,366,445.
How plausible does it sound to you? Well, never mind. The point is that
we have just used an utterance interpretively, to represent an assumption,
without attributing this assumption to anyone: that is, without reporting
it. We have already done so many times in this book: many of our
numbered examples are used to represent utterances, assumptions or
intentions which we did not attribute to anybody, not even to fictitious
characters, and which we put forward to illustrate some abstract point.
In speculative thinking, thoughts are often entertained as approximate
representations of assumptions one would like to be able to formulate
better. This is true of trivial speculation: I don't remember when the party
at the Jones's is supposed to be; I tryout on myself, 'It's on Tuesday', 'It's
on Wednesday', 'It's on Thursday', ete., hoping that when I hit on the
right date, I will somehow recognise it. I entertain these successive
thoughts as attempts at representing the relevant piece of information in
my memory, and this is what makes my hope not entirely unreasonable: a
230
mental match may occur, whereas if these thoughts were to be processed

as descriptions, I would have to wait until the event took place to verify
one of them and falsify the others. In scientific speculation too, inadequate
or incomplete hypotheses are knowingly entertained, not as descriptions
of the empirical phenomena under study, but as temporary representations of better hypotheses to come.
Reported speech or thought is thus not the only interpretive use of
language. Utterances can be used interpretively to represent utterance
types or thoughts which are worth considering for their intrinsic
properties, and not because they can be attributed to Peter, Mary, the
inn-keeper, or public opinion. But we want to argue that there is an even
more essential interpretive use of utterances: on a more fundamental level,
every utterance is used to represent a thought of the speaker's.
One of the assumptions a speaker intends to make manifest is that she is
entertaining some thought with some particular attitude, since it is on this
ground that the hearer may be led to entertain a similar thought with a
similar attitude. You may well tell me that you will come tomorrow: you
will not make me believe it unless you first make me believe that you
believe it too. That much is hardly controversial. Actually, an even
stronger claim is generally made. Most pragmatists and philosophers of
language take for granted that there is a convention, principle or
presumption/ 3 to the effect that the meaning of the utterance must be a
literal expression, i.e. an identical reproduction, of a thought of the
speaker's. We believe that this claim is too strong. People certainly do not
express themselves literally all the time, and when they do not, there is no
intuition that a norm has been transgressed. There is thus no empirical
evidence for a convention of literalness or anything of the sort. Such a
convention is postulated on purely theoretical grounds: the underlying
code model of communication implies that utterances are understood as
communicating what they encode; then non-literal uses can be analysed as
more or less codified departures from literalness, recoverable by inference.
Our approach is different: we have rejected the code model and we hope
to explain how verbal communication is possible without postulating any
ad hoc constraint apart from strictly grammatical constraints. Code
theorists see verbal communication as involving a speaker encoding one of
her thoughts in an utterance, which is then decoded by the hearer (with an
extra layer of inference in modern versions). We see verbal communication as involving a speaker producing an utterance as a public interpretation of one of her thoughts, and the hearer constructing a mental
interpretation of this utterance, and hence of the original thought. Let us
say that an utterance is an interpretive expression of a thought of the
speaker's, and that the hearer makes an interpretive assumption about the
231
speaker's informative intention. It follows from our general account of

inferential communication that an utterance should be an interpretive
expression of a thought of the speaker's. However, we see no reason to
postulate a convention, presumption, maxim or rule of literalness to the
effect that this interpretation must be a literal reproduction. How close the
interpretation is, and in particular when it is literal, can be determined on
the basis of the principle of relevance.
We assume, then, that every utterance is an interpretive expression of a
thought of the speaker's. What does that thought itself represent, and
how? A mental representation, like any representation with a propositional form, can be used descriptively or interpretively. When it is used
descriptively, it can be a description of a state of affairs in the actual
world/4 or it can be a description of a desirable state of affairs. When it is
used interpretively, it can be an interpretation of some attributed thought
or utterance, or it can be an interpretation of some thought which it is or
would be desirable to entertain in a certain way: as knowledge, for
instance. There may be other possibilities, and one might consider what
the thoughts interpreted by thoughts might represent in their turn and
how, but let us leave it at that, and use figure 3 (p.232) to show the
representations and relationships considered so far.
Any utterance involves at least two relationships: a relationship
between its propositional form and a thought of the speaker's, and one of
the four possible relationships between that thought and what it
represents. All the basic relationships involved in tropes and illocutionary
forces are represented in this diagram, as we will show in the next three
sections. Our argument may be summarised as follows: metaphor
involves an interpretive relation between the propositional form of an
utterance and the thought it represents; irony involves an interpretive
relation between the speaker's thought and attributed thoughts or
utterances; assertion involves a descriptive relation between the speaker's
thought and a state of affairs in the world; requesting or advising involves
a descriptive relation between the speaker's thought and a desirable state
of affairs; interrogatives and exclamatives involve an interpretive relation
between the speaker's thought and desirable thoughts. These claims will
be discussed in more detail below.
8 Literalness and metaphor
In this section we will consider the relationship at the top of figure 3:

between the propositional form of an utterance and the thought this
utterance is used to represent. We have argued that in general, the
relationship is one of resemblance rather than identity between proposi-
232
The propositional form

of an utterance
is
an interpretation of
a mental representation
of the speaker
which can be
entertained as
a description of
an interpretion of
an actual
(e.g. attributed)
representation
a desirable
(e.g. relevant)
representation
an actual
state of affai rs
a desirable
state of affairs
(a)
(b)
(c)
(d)
Figure 3
tional forms. We are treating literalness, or identity of propositional

forms, as a limiting case rather than a norm. We will show that this
approach, combined with relevance theory, yields a straightforward
account of metaphors and related tropes.
It might be thought that in even contemplating a move in this direction
we are venturing onto dangerous ground. 'Resemblance' is notoriously
not a well-defined term. Anything may resemble anything in at least some
respect. When and how a resemblance is perceived is an open question in
cognitive psychology; the mechanisms involved are ill understood.
However, since we suspect that a proper account of the perception of
resemblance in general should be based on a well-developed notion of
233
relevance, we do not feel too dismayed. Moreover, for the time being we
are concerned only with resemblances of a very restricted type: logical
resemblances among propositional forms (where two propositional forms
resemble each other if and only if they share logical properties ). We will
show that the identification of these resemblances, like every other aspect
of comprehension, is guided by the principle of relevance.
Let us say that an utterance, in its role as an interpretive expression of a
speaker's thought, is strictly literal if it has the same propositional form as
that thought. To say that an utterance is less than strictly literal is to say
that its propositional form shares some, but not all, of its logical properties
with the propositional form of the thought it is being used to interpret.
From the standpoint of relevance theory, there is no reason to think that
the optimally relevant interpretive expression of a thought is always the
most literal one. The speaker is presumed to aim at optimal relevance, not
at literal truth. The optimal interpretive expression of a thought should
give the hearer information about that thought which is relevant enough
to be worth processing, and should require as little processing effort as
possible. There are many quite ordinary situations where a literal
utterance is not optimally relevant: for example, where the effort needed
to process it is not offset by the gain in information conveyed. There are
thus many situations where a speaker aiming at optimal relevance should
not give a literal interpretation of her thought, and where the hearer
should not treat her utterance as literal.
For example, suppose I earn 797.32 pence a month. You, a friend I
have not seen for some years, ask me over a drink how much I am earning
now. If I remember the exact figure, I can choose between the strictly
literal and truthful answer in (1 OSa), and the less than literal (1 OSb), which
I know to he strictly speaking false:
(105) (a) I earn 797.32 pence a month.
(h) I earn 800 a month.
In the circumstances, there is no reason to think you need an exact figure.
From either reply you will be able to derive exactly the same conclusions
about my status, standard of living, purchasing power, life style, and
whatever else you are planning to use my salary as an indicator of. Aiming
at optimal relevance, I should therefore choose the reply which will
convey these conclusions as economically as possible. In other words, I
should choose the false hut economical (1 OSh) rather than the complex but
strictly literal and truthful (10Sa), and expect you to recognise that I am
offering something less than a strictly literal interpretation of my
thoughts.
To take a rather more abstract example, suppose I have a complex
thought P, which makes manifest to me a set of assumptions I, and I

want to communicate I to you. Now suppose that the following
conditions are met: P is too complex to be represented literally, but the
assumptions in I are all straightforwardly derivable as logical or
contextual implications of an easily expressed assumption Q. The
problem is that Q is not a thought of mine; it has some logical and
contextual implications which I do not accept as true and which I do not
want to communicate. What should I do? Given the principle of relevance,
as long as you have some way of sorting the implications of Q into those I
do and those I do not want to endorse, the best way of communicating I
may well be to express the single assumpti'on Q and leave the sorting to
In these circumstances, the utterance which expresses Q is an interpretive expression of my complex thought P: they share logical properties,
more specifically logical and contextual implications. Moreover, the
criterion of consistency with the principle of relevance provides a means
of distinguishing those contextual implications which are shared from
those which are not; that is, it gives you a way of constructing the right
interpretive assumption about my informative intention.
We are assuming that all the hearer can take for granted is that an
utterance is intended as an interpretation of one of the speaker's thoughts.
This does not mean that whenever an assumption is expressed, the hearer
has to compute all its logical and contextual implications and sort through
them one by one to find out which subset of them are implications of the
speaker's thought. In the framework we are proposing, this wasteful
manoeuvre is quite unnecessary. If the speaker has done her job correctly,
all the hearer has to do is start computing, in order of accessibility, those
implications which might be relevant to him, and continue to add them to
the overall interpretation of the utterance until it is relevant enough to be
consistent with the principle of relevance. At this point, the sorting will
have been accomplished as a by-product of the search for relevance, and
will require no specific effort of its own.
It follows that the hearer should take an utterance as fully literal only
when nothing less than full literality will confirm the presumption of
relevance. In general, some looseness of expression is to be expected. For
example, if someone says, 'It's 5 p.m.', she should not be taken to task if it
turns out to be five minutes or two minutes to, unless the relevance of the
utterance depends on that kind of exactitude. If someone says 'I'm
exhausted', there is no point in quibbling over whether exhausted is
exactly what she is: as long as she can be taken to have conveyed an
acceptable range of implications, she will have achieved optimal relevance.
The examples discussed so far would normally be treated as loose uses
of language, but would not be regarded as figurative: there is no
temptation to invoke the substitution of a figurative for a literal meaning.
234
235
We want to claim that there is no discontinuity between these loose uses

and a variety of 'figurative' examples which include the most characteristic
examples of poetic metaphor. In both cases, the propositional form of the
utterance differs from that of the thought interpreted. In both cases, the
hearer can proceed on the assumption that these two propositional forms
have some identifiable logical and contextual implications in common. In
both cases, the same interpretive abilities and procedures are involved.
Consider first an example of hyperbole. The speaker expresses, but does
not explicate, the assumption in (1 06a), and implicates the weaker (1 06b):
(106) (a) Bill is the nicest person there is.
(b) Bill is a very nice person.
How can this be consistent with the principle of relevance? Let us assume
that by explicating (106b) directly the speaker would not exhaust her
thoughts about Bill: its contextual effects would fall short of what she
wants to convey. Nor is there any obvious combination of adverbs and
adjectives that would exactly express her thoughts. Perhaps they are too
vague: there are a lot of aspects of Bill's niceness that she is not thinking
about with equal clarity at the time, and to access these thoughts and make
them more precise would involve more work than she is prepared to do.
She can be sure, on the other hand, that all the assumptions she wants to
communicate are among the logical and contextual implications of (106a).
(106a) has other implications which she does not want to communicate.
As long as she can rely on the hearer to ignore or discard them, (106a) will
be a much more adequate interpretation of her thoughts than the weaker
(106b).
What exactly does (106a) convey? The speaker is certainly strongly
implicating (106b). However, if this were all she had wanted to convey,
she could have saved the hearer some processing effort by expressing
(106b) directly. As always, the element of indirectness in an utterance
must be offset by some increase in contextual effects. By expressing
(106a), the speaker thus encourages the hearer to look for a range of
further contextual implications not shared, or not equally strengthened,
by (1 06b), and assume that within this range there are some she intends to
implicate. He might thus begin running through the names of their
common acquaintances and conclude that the speaker prefers Bill to each
of these; he might conclude that Bill has behaved in ways that are so
admirable that the speaker can find no words to describe them, and so on.
The wider the range of possible conclusions, the weaker the implicatures,
and the more the hearer must share the responsibility for deriving them.
Thus (106a) conveys, on the one hand, a suggestion that the speaker has a
certain attitude to Bill, a certain vision of Bill and his niceness, and on the
236

other hand, an incentive to the hearer to develop his own vision of Bill and
conclude that it overlaps to some extent with hers.
Let us return, then, to our original example of metaphor:
(88) This room is a pigsty.
This is a very standardised metaphor. Typically, such examples give
access to an encyclopaedic schema with one or two dominant and highly
accessible assumptions. Thus pigsties are stereotypically filthy and
untidy. When (88) is processed in this stereotypical context, it will yield
the implication that the room is filthy and u:ntidy. If the speaker had not
intended this implication to be derived, she should have rephrased her
utterance to eliminate it: hence (88) strongly implicates that the room is
filthy and untidy. However, the speaker must have intended to convey
something more than this if the relative indirectness of the utterance is to
be justified: an image, say, of filthiness and untidiness beyond the norm,
beyond what could have been satisfactorily conveyed by saying merely
'This room is very filthy and untidy.' Thus even this very standardised
example cannot be paraphrased without loss.
Moving to a marginally more creative example, (107) is a fairly
conventional metaphor whose interpretation involves bringing together
the encyclopaedic entries for Robert and bulldozer, which do not
normally come together in a subject-predicate relationship:
(107) Robert is a bulldozer.
The result will be a wide array of contextual implications, many of which,
being contradictory, can be automatically discarded. The relevance of
(107) will be established by finding a range of contextual effects which can
be retained as weak or strong implicatures. Here there is no single strong
implicature that automatically comes to mind, but rather a slightly
weaker, less determinate range having to do with Robert's persistence,
obstinacy, insensitivity and refusal to be deflected. The hearer thus has to
take a slightly greater responsibility for the resulting interpretation than
he does with (106a) and (88).
In general, the wider the range of potential implicatures and the greater
the hearer's responsibility for constructing them, the more poetic the
effect, the more creative the metaphor. A good creative metaphor is
precisely one in which a variety of contextual effects can be retained and
understood as weakly implicated by the speaker. In the richest and most
successful cases, the hearer or reader can go beyond just exploring the
immediate context and the entries for concepts involved in it, accessing a
wide area of knowledge, adding metaphors of his own as interpretations of
possible developments he is not ready to go into, and getting more and
more very weak implicatures, with suggestions for still further processing.
237
The result is a quite complex picture, for which the hearer has to take a
large part of the responsibility, but the discovery of which has been
triggered by the writer. The surprise or beauty of a successful creative
metaphor lies in this condensation, in the fact that a single expression
which has itself been loosely used will determine a very wide range of
acceptable weak implicatures.
Take, for example, Flaubert's comment on the poet Leconte de Lisle:
(108) His ink is pale. (Son encre est pale.)
A strictly literal construal of this utterance is clearly ruled out: it is hard to

see what relevance could attach to knowing the colour of a poet's ink. Nor
is there any obvious strong implicature. The only way of establishing the
relevance of this utterance is to look for a wide range of very weak
implicatures. This requires several extensions of the context. In the most
accessible context of encyclopaedic information about ink and handwriting, most implications are irrelevant: after all, Leconte de Lisle's poetry is
read not in his handwriting but in print; the only clear implicature in this
first context is that he has the character of a man who would use pale ink.
Some other implications - that Leconte de Lisle's writing lacks contrasts,
that it may fade - have further relevant implications in a context to which
has been added the premise that what is true of his handwriting is true of
his style. Someone who knows little of Leconte de Lisle's work might
conclude, for example, that there is something weak about his poetry, that
his writings will not last, that he does not put his whole heart into his
work, and so on. Someone who has a deeper acquaintance with the poet
would be able to construe the criticism in much more detailed and pointed
ways. The resulting interpretation, with its characteristic poetic effect,
owes a lot simultaneously to Flaubert, for foreseeing how it might go, and
to the reader, for actually constructing it.
On this approach, metaphor and a variety of related tropes (e.g.
hyperbole, metonymy, synecdoche) are simply creative expioitations of a
perfectly general dimension of language use. The search for optimal
relevance leads the speaker to adopt, on different occasions, a more or a
less faithful interpretation of her thoughts. The result in some cases is
literalness, in others metaphor. Metaphor thus requires no special
interpretive abilities or procedures: it is a natural outcome of some very
general abilities and procedures used in verbal communication. In the next
section, we will show that the same is true of irony.
9 Echoic utterances and irony

We would now like to show that irony and a variety of related tropes (e.g.
238
meiosis, litotes) fall together with a range of cases which would not
normally be regarded as figurative at all. What unites these cases is the fact
that the thought of the speaker which is interpreted by the utterance is
itself an interpretation. It is an interpretation of a thought of someone
other than the speaker (or of the speaker in the past). That is, these
utterances are second-degree interpretations of someone else's thought, as
illustrated by path (a) in figure 3 above. If we are right, then the same is
true of irony as is true of metaphor: whatever abilities and procedures are
needed to understand it are independently needed for the interpretation of
quite ordinary non-figurative utterances. 25
We have already considered, in section 7, the case of utterances used to
interpret someone else's speech or thought. They are always (at least)
second-degree interpretations: like all utterances, they first interpret a
thought of the speaker, and it is only because this thought is itself an
interpretation of someone else's thought that the utterance ultimately
represents someone else's thought. Another way of making the same point
is to say that an utterance used as an interpretation of someone else's
thought is always, in the first place, an interpretation of one's understanding of that other person's thought. When we talk of utterances used to
interpret someone else's thought, it should be clear, then, that we are
always talking of second-degree interpretations.
How do interpretations of someone else's thought achieve relevance? In
the best-known case, that of 'reported speech', they achieve relevance by
informing the hearer of the fact that so-and-so has said something or
thinks something. In other cases, these interpretations achieve relevance
by informing the hearer of the fact that the speaker has in mind what
so-and-so said, and has a certain attitude to it: the speaker's interpretation
of so-and-so's thought is relevant in itself. When interpretations achieve
relevance in this way, we will say that they are echoic, and we will argue
that ironical utterances are cases of echoic interpretation.
Here is a simple case of an echoic utterance:
(109) Peter: The Joneses aren't coming to the party.
Mary: They aren't coming, hum. If that's true, we might invite the
Smiths.
Mary's first sentence echoes what Peter has just said. It achieves relevance
not, of course, by reporting to Peter what he has just said, but by giving
evidence that Mary has paid attention to his utterance and is weighing up
its reliability and implications.
An echoic utterance need not interpret a precisely attributable thought:
it may echo the thought of a certain kind of person, or of people in general.
Suppose you tell me to hurry up and I reply as follows:
239
(110) More haste, less speed.

This utterance is a literal interpretation of a traditional piece of wisdom
which achieves relevance by making manifest that I find this piece of
wisdom indeed wise in the circumstances. Clearly, however, what makes
traditional wisdom traditional is that it is attributable not to any specific
source but to people in general.
By representing someone's utterance, or the opinions of a certain type
of person, or popular wisdom, in a manifestly sceptical, amused,
surprised, triumphant, approving or reproving way, the speaker can
express her own attitude to the thought echoed, and the relevance of her
utterance might depend largely on this expression of attitude. Sometimes,
the speaker's attitude is left implicit, to be gathered only from tone of
voice, context and other paralinguistic clues; at other times it may be
made explicit. We will argue that verbal irony invariably involves the
implicit expression of an attitude, and that the relevance of an ironical
utterance invariably depends, at least in part, on the information it
conveys about the speaker's attitude to the opinion echoed.
There is no limit to the attitudes that a speaker can express to an opinion
echoed. In particular, she may indicate her agreement or disagreement
with it. Compare (111) and (112):
(111) (a) Peter: It's a lovely day for a picnic.
[They go for a picnic and the sun shines.]
(b) Mary (happily): It's a lovely day for a picnic, indeed.
(112) (a) Peter: It's a lovely day for a picnic.
[They go for a picnic and it rains.]
(b) Mary (sarcastically): It's a lovely day for a picnic, indeed.
In both (111 b) and (112b) there is an echoic allusion to be picked up. In the
circumstances described, it is clear that the speaker of (111 b) endorses the
opinion echoed, whereas the speaker of (112b) rejects it with scorn. These
utterances are interpreted on exactly similar patterns; the only difference
is in the attitudes they express. (111 b) has not been thought by
rhetoricians to be worthy of special attention; (112b) is, of course, a case of
verbal irony.
The attitude expressed by an ironical utterance is invariably of the
rejecting or disapproving kind. The speaker dissociates herself from the
opinion echoed and indicates that she does not hold it herself. Indeed, it
may be obvious in the circumstances that she believes the opposite of the
opinion echoed: thus the speaker of (112b) manifestly believes that it is not
a lovely day for a picnic. From this it follows that it was wrong of her
companion to say that it was a lovely day for a picnic, that his judgement
240
has been unsound, that they should never have set out, that it was his fault
that their day has been ruined, and so on. The recovery of these
implicatures depends, first, on a recognition of the utterance as echoic;
second, on an identification of the source of the opinion echoed; and third,
on a recognition that the speaker's attitude to the opinion echoed is one of
rejection or dissociation. We would argue that these are common factors
in the interpretation of all ironical utterances.
As regards the particular range of rejecting or dissociative attitudes
conveyed by verbal irony, there is no need to look for a clear-cut answer.
Are anger, outrage and irritation among the ~ttitudes that the ironist can
convey? This question, it seems to us, should be of interest only to
lexicographers. From the pragmatic point of view, what is important is
that a speaker can use an echoic utterance to convey a whole range of
attitudes and emotions, ranging from outright acceptance and endorsement
to outright rejection and dissociation, and that the recognition of these
attitudes and emotions may be crucial to the interpretation process. We
doubt very much that there is either a well-defined subset of ironical
attitudes or a well-defined subset of ironical utterances which express
them. Rather, what exists is a continuum, with different blends of attitude
and emotion giving rise to a whole range of borderline cases which do not
fit neatly into any existing scheme. Irony is not a natural kind.
Let us briefly compare this account with the classical account of irony as
saying one thing and meaning, or implicating, the opposite. The most
obvious problem with the classical account- and with its modern variant,
the Gricean account - is that it does not explain why a speaker who could,
by hypothesis, have expressed her intended message directly should
decide instead to say the opposite of what she meant. It cannot be too
strongly emphasised what a bizarre practice this would be. Suppose we are
out for a drive and you stop to look both ways before joining the main
road. The road is clear, but as you are about to drive on I say quietly,
(113) There's something coming.
You slam on your brakes and look both ways, but the road is as clear as
before. When you ask me what on earth I was doing, I explain gently that I
was merely trying to reassure you that the road was clear. My utterance
satisfies the classical definition of irony. I have said something which is
patently false, and there is a logically related assumption, namely (114),
which I could truthfully have expressed:
(114) There's nothing coming.
Why do you not instantly leap to the conclusion that this is what I was
trying to convey?
The classical account of irony notably fails to explain what distinguishes
241
genuine irony from the mere irrationality exhibited by (113). In our

framework, the difference is clear. Genuine irony is echoic, and is
primarily designed to ridicule the opinion echoed. Let us recast our
example so that these conditions are satisfied. You are an over-cautious
driver, constantly on the alert for danger, who never pulls into a main road
in front of oncoming traffic, however far away. When we stop at the
junction, the road is straight and entirely clear in both directions except
for a just-visible cyclist on the horizon. As you pull onto the main road, I
say, reprovingly, (113). In the circumstances, this remark may well be
ironical. I am echoing back to you the sort of opinion you are constantly
expressing, but in circumstances which make it clearly ridiculous. Thus,
all that is needed to make (113) ironical is an echoic element and an
associated attitude of mockery or rejection.
Notice how inadequate it would be with this example to say that I was
merely trying to implicate the opposite of what I had said. (114) is at most
an implicated premise of my utterance, and certainly does not constitute
its main point. The main point of the utterance is to express my attitude to
the sort of opinions you are constantly expressing, and in doing so to
implicate that you are over-cautious, that you are making a fool of
yourself by worrying, and so on. If I had merely wanted to convey (114) I
would of course have expressed this assumption directly.
In fact there are many examples of irony which fall outside the scope of
the classical definition of irony as saying one thing and meaning the
opposite. Consider (llsa), cited as an example of irony in many of the
standard works: c
(115) (a) When all was over and the rival kings were celebrating their
victory with Te Deums in their respective camps ... (Voltaire,
Candide)
To treat this utterance as equivalent to (l1sb) or (l1sc) would be not:
merely lame but positively wrong:
(115) (b) When all was over and the rival kings were not celebrating their
victory with Te Deums in their respective camps ...
(c) When all was over and the rival kings were bewailing their defeat:
with Misereres in their respective camps ...
Voltaire was not suggesting that neither side won the battle and celebrated
victory, nor that both sides lost and bewailed their defeat. The point of the
utterance lies elsewhere. Our framework explains both the intuition that
this is a genuine case of irony and the fact that it does not implicate (llsb)
or (l1sc). Voltaire is echoing claims made by the rival kings. Since the
claims contradict each other, it is clear that if he is minimally alert he
cannot be endorsing them both: that indeed he must believe, and expect
242
his audience to believe, that at least one of them is false. However, there is
no need to come to the stronger conclusion that there is some determinate
assumption which means the opposite of what has been explicitly said, and
which Voltaire wanted to endorse.
In fact (l1Sa), like many of the best examples of irony, is a garden-path
utterance, likely to cause the reader momentary processing difficulties
later offset by appropriate rewards. One at first reads it as an ordinary
assertion, is led to the absurd conclusion that both sides won, and only
then reinterprets echoically. By leaving the echo implicit when the
addition of some explicit material would have:immediately put the reader
on the right track, the author opens up a whole new line of interpretation.
What sort of hearer would have needed no explicit push towards the
echoic interpretation? One who would automatically assume that after a
battle both sides invariably claim victory, that this behaviour is always
absurd, that the author and reader are not the sort of people to be fooled,
and so on. Thus, by leaving the echo implicit, the author manages to
suggest that he shares with his readers a whole cynical vision which is
absent from the explicitly interpretive version in (llSd):
(115) (d) When the battle was over and the rival kings were doing what
they described as celebrating their victory with Te Deums in
their respective camps ...
Example (86) in section 7 fits quite straightforwardly into this
framework.
(86) Peter is quite well-read. He's even heard of Shakespeare.
To believe (86), one would also have to believe that anyone who has heard
of Shakespeare is quite well-read - a patently ludicrous opinion. The
speaker of (86) thus makes fun of the idea that Peter is well-read, and
strongly implicates that he is not well-read at all. However, the irony
would fall flat if, manifestly, neither Peter himself nor anyone else had
ever entertained the thought that Peter was well-read: in this case there
would be no one to echo.
Our accounts of metaphor and irony share two essential features. First,
we are arguing that the possibility of expressing oneself metaphorically or
ironically and being understood as doing so follows from very general
mechanisms of verbal communication rather than from some extra level of
competence. 26 Second, we are arguing that there is a continuum of cases
rather than a dividing line between metaphorical and literal utterances on
the one hand, between ironical utterances and other echoic utterances on
the other; we are arguing, in other words, that metaphor and irony involve
no departure from a norm, no transgression of a rule, convention or
maXim.
Speech acts
243
If our account is correct, there are two conclusions to be drawn: first,
metaphor and irony are not essentially different from other types of
'non-figurative' utterances; and second, they are not essentially similar to
one another. Metaphor plays on the relationship between the propositional form of an utterance and the speaker's thought; irony plays on the
relationship between the speaker's thought and a thought of someone
other than the speaker. This suggests that the notion of a trope, which
covers metaphor and irony and radically distinguishes them from
'non-figurative' utterances, should be abandoned altogether: it groups
together phenomena which are not closely related and fails to group
together phenomena which are.
10 Speech acts
Perhaps the single most uncontroversial assumption of modern pragmatics is that any adequate account of utterance comprehension must include
some version of speech-act theory. As Levinson (1983: 226) says,
speech acts remain, along with presupposition and implicature in
particular, one of the central phenomena that any general pragmatic
theory must account for.
We would like to question this assumption. The vast range of data that
speech-act theorists have been concerned with is of no special interest to
pragmatics. What is of interest is their attempt to deal with the
interpretation of non-declarative (e.g. interrogative and imperative)
sentences, which must indeed be accounted for in any complete pragmatic
theory. In this section we will look first at speech-act theory as a general
pragmatic programme, and then at the analysis of non-declaratives, for
which we will sketch some proposals of our own.
Speech-act theory grew out of a reaction to what was seen as an
excessively narrow concentration on the informative use of language.
Language can be used to perform actions - speech acts: for example, to
create and discharge obligations, to influence the thoughts and actions of
others, and more generally, to create new states of affairs and new social
relationships. A better understanding of the nature of language, argued
Austin (1962), must involve a better understanding of how language is
embedded in social institutions, and of the various actions that it can be
used to perform.
Speech-act theorists have been much concerned with descriptive
questions: how many types of speech act are there, and how should they
be grouped together ?27 Searle (1979a) distinguishes assertives (e.g. state-

ments), which commit the speaker to the truth of the assumption
expressed; directives (e.g. orders), attempts to get the hearer to do
something; commissives (e.g. promises), which commit the speaker to the
performance of a future action; expressives (e.g. congratulations), which
convey the speaker's emotional attitude to the assumption expressed; and
declarations (e.g. declaring the court open), which bring about the state of
affairs described in the assumption expressed.
However, there have also been some explanatory attempts to show how
utterances are assigned to speech-act types and how indirect or implicit
speech acts are performed. The 'recognition;: of indirect speech acts is
generally seen as proceeding along Gricean lines. Consider (116), for
example:
244
(116) The battery's gone flat.

This might be analysed as a direct assertion that the battery had gone flat.
It is easy to think of circumstances in which a speaker who asserted (116)
would also implicate (117) or (118):
(117) The hearer shouldn't have let the battery go flat.
(118) The hearer should get the battery recharged.
According to speech-act theory, these implicatures too should be assigned

to speech-act types: thus, (117) might be analysed as an accusation or
reproof, and (118) as a request or an order. Speech-act theory thus offers
itself as a natural complement to Gricean pragmatics, dealing with the
classification in speech-act terms of both explicatures and implicatures.
A crucial assumption behind this pragmatic programme is that the
assignment of every utterance to a particular speech-act type is part of
what is communicated and plays a necessary role in comprehension. What
is surprising is how little attention has been paid to justifying this
assumption. It is one thing to invent, for one's own theoretical purposes, a
set of categories to use in classifying the utterances of native speakers, or to
try to discover the set of categories that native speakers use in classifying
their own utterances. It is quite another to claim that such a classification
plays a necessary role in communication and comprehension. To see the
one type of investigation as necessarily shedding light on the other is
rather like moving from the observation that tennis players can generally
classify strokes as volleys, lobs, approach shots, cross-court backhands
and so on, to the conclusion that they are unable to perform or return a
stroke without correctly classifying it. The move clearly requires some
justification.
Some speech acts do have to be communicated and identified as such in
order to be performed. Bidding two no trumps at bridge is an example. In
Speech acts
245
order to perform this speech act, the speaker must ostensively communicate, by linguistic means, via an utterance such as (119a), or by inference,
via an utterance such as (119b), an assumption of the form in (120):
(119) (a) I bid two no trumps.
(b) Two no trumps.
(120) The speaker is bidding two no trumps.
However, the study of bidding is part of the study of bridge, not of verbal
communication. Generally speaking, the study of institutional speech acts
such as bidding, or declaring war, belongs to the study of institutions.
Many other speech acts, by contrast, can be successfully performed
without being identified as such either by the speaker or by the hearer.
Take predicting, for example. What makes an utterance a prediction is not
the fact that the speaker ostensively communicates that she is making a
prediction; it is that she ostensively communicates an assumption with a
certain property: that of being about a future event at least partly beyond
her control. Thus, (121) could be a prediction without the speaker's ever
intending to communicate, or the hearer's ever recovering, the information in (122):
(121) The weather will be warmer tomorrow.
(122) The speaker is predicting that the weather will be warmer tomorrow.
This is not to say that it would never be desirable for the speaker of (121)
simultaneously to communicate assumption (122), or that it would never
be relevant for the hearer of (121) to recognise it as a prediction. The fact
that a prediction is being made is a fact like any other, and as such may be
made manifest by a speaker, or recognised by a hearer, in the usual way.
Our claim is simply that even where (122) is manifestly true, its recovery is
not essential to the comprehension of an utterance such as (121), as the
recovery of (120) is essential to the comprehension of an utterance such as
(119b) above.
Many speech acts which have been regarded as quite central to
pragmatics fall into one or other of these two categories. Promising and
thanking, for example, fall into the first category: they are institutional
acts, which can be performed only in a society with the requisite
institutions, and which must be recognised as such in order to be
successfully performed. 28 By contrast, asserting, hypothesising, suggesting, claiming, denying, entreating, demanding, warning and threatening
(to the extent that they are speech acts at all) fall into the second category:
they are acts which do not need to be identified as such in order to be
successfully performed, and which, like predicting, can be identified in
terms of some condition on their explicit content or implicatures. In
246
neither case does the interpretation of utterances involving such speech

acts require any special pragmatic principles or machinery not already
needed on independent grounds.
There is, however, a small class of speech acts which fall into neither of
these categories, and which are of genuine interest to pragmatics. They
include saying, telling and asking. Consider (123)-(125):
(123) You will finish the work before 6 p.m.
(124) Will you finish the work before 6 p.m.?
(125) Finish the work before 6 p.m.
It is clear that a declarative such as (123), an interrogative such as (124) and

an imperative such as (125) exhibit both logical similarities and differences. Their similarities can be accounted for by assuming that they have
the same or similar logical forms. Speech-act theory seems to offer a way
of accounting for their differences. It is often suggested, for example, that
there are systematic correlations between syntactic sentence type and
speech-act type, so that a declarative such as (123) is correlated with the
speech-act of saying that the hearer will finish the work before 6 p.m., an
interrogative such as (124) with the speech act of asking whether the hearer
will finish the work before 6 p.m., and an imperative such as (125) with the
speech act of telling the hearer to finish the work before 6 p.m. We
adopted something like this assumption in earlier chapters, in suggesting
that the propositional form P of an ordinary assertion is standardly
integrated into an assumption schema of the form The speaker said that P.
If we are right, the recovery of such descriptions is an essential part of
the comprehension process, and the speech acts of saying, asking and
telling do not fall into our second category of speech acts. However, it is
also clear that the acts of saying, asking and telling are neither social nor
institutional in the way that bidding at bridge, promising and thanking
are. It is easy to think of societies which lack the institution of bridge; we
would also maintain that there are societies which lack the institutions of
promising and thanking. Saying, telling and asking, by contrast, are
universal, and appear to be genuinely communicative rather than socialinstitutional categories.
However, to say that these three generic speech acts have a role to play
in pragmatic theory is not to say that a theoretically adequate account of
them already exists. It is tempting to assume that saying that is simply the
most general type of assertive speech act, telling to is simply a general,
action-requesting type of directive, and asking whether is simply a
general, information-requesting type of directive. However, if the
correlation between syntactic sentence type and speech-act type is to be
maintained, saying that cannot be a type of assertive at all. An assertive is a
speech act which commits the speaker to the truth of the propositional
Speech acts
247
form of her utterance; but as we have seen, not all declarative utterances
are assertive in this sense: for example, metaphors and ironies are not. The
problem is quite general. If a directive is an attempt to get the hearer to
perform the action explicitly described, then the ironical imperative (126)
is not a directive:
(126) Go ahead and ruin my carpet.
It is not a genuine attempt to get the hearer to go ahead and ruin the
speaker's carpet. Similarly, the rhetorical question (127) is not a genuine
request for information:
(127) What monster would dare to harm a sleeping child?
Thus, the correlation between syntactic sentence types and generic speech
acts cannot be maintained unless a whole range of utterance types such as
(126) and (127) are excluded as 'insincere' or 'defective', or the traditional
typology of speech-act types is abandoned.
Even the claim that there is a well-defined range of mutually exclusive
syntactic sentence types is open to question. Is (128), which can be used
with either assertive or directive force, a declarative or an imperative?
(128) You are to leave tomorrow.
Is (129), said with rising intonation, a declarative or an interrogative?
(129) You won't be needing the car?
Is (130) a declarative or an exclamative?
(130) This book is so interesting.
e
a
Ll
What undeniably exists is not a well-defined range of syntactic sentence

types but a variety of overt linguistic devices - e.g. indicative, imperative
or subjunctive mood, rising or falling intonation, inverted or uninverted
word order, the presence or absence of Wh-words, or of markers such as
'let's' or 'please' - which can guide the interpretation process in various
ways. While it may be possible to build a theory of syntactic sentence
types around these devices, as far as we know this work has not yet been
done. In what follows, the use of such terms as 'declarative sentence',
'interrogative sentence' and so on should be regarded as nothing more
than a convenient shorthand. 29
Let us define saying that P, where P is the propositional form of the
utterance, as communicating that the thought interpreted by P is
entertained as a description of an actual state of affairs. It may be
entertained as a true description by the speaker, or by the person or type of
person whose thought is being interpreted in the second degree. When
you say that P, you communicate that you are saying that P. You may
248
communicate this by means of linguistic indicators such as indicative

mood, declarative word order and so on; in the absence of such linguistic
indicators, as in telegraphic forms of writing or speech, it is up to the
hearer to decide whether the speaker is saying that P or performing one of
the other generic speech acts. In this, as in every other aspect of
interpretation, he should adopt the first assumption that is consistent with
the principle of relevance.
On hearing (131), the speaker should identify the propositional form of
the utterance and integrate it into the description in (132):
(131) The bus is leaving.
(132) The speaker has said that the bus is leaving.
As we have seen, this description can be relevant in a variety of ways. For
example, it might provide the hearer with evidence for (133), and if he
trusts the speaker enough, for (134):
(133) The speaker believes that the bus is leaving.
An utterance which is intended to achieve relevance in this way is, of
course, an ordinary assertion. Ordinary assertions are the result of
choosing path (c) on figure 3 above, and producing an utterance which is a
fully literal interpretation of the speaker's thought.
Utterance (131) could be metaphorical: imagine that it is said, with no
bus in sight, to someone hesitating about joining a group of friends who
are all ready to go for a walk and are waiting for his decision. In that case
(132) would make manifest (135) and, provided that the hearers trust the
speaker enough, (136), where (136) is a contextual implication which (131)
manifestly shares with the thought it is used to interpret:
(135) The speaker believes that if the hearer does not decide to go
immediately, it will be too late.
(136) If the hearer does not decide to go immediately, it will be too late.
A metaphorical utterance of this type would be the result of choosing path
(c) on figure 3 above, and producing an utterance which is a less than fully
literal interpretation of the speaker's thought.
Or (131) might be put forward as a report of speech: say a report of what
the bus driver has just said. In that case, (132) might provide the hearer
with evidence for (137) and, if he trusts the speaker and the bus driver
enough, for (138) and (139):
(137) The speaker believes the bus driver has said that the bus is leaving.
(138) The bus driver has said that the bus is leaving.
Speech acts
249
In this case the utterance would be the result of choosing path (a) on figure
3 above.
As, ,:,e ~av~ seen, ~ome acts of saying that P ~chieve relevance not by
prov1dlllg llldIrect eVIdence for P but by expresslllg the speaker's attitude
to P. For example, the speaker of (131) above, in reporting the bus driver's
words, may tacitly dissociate herself from them. In this case, (132) might
achieve relevance by providing the hearer with evidence for (140) and, if he
trusts the speaker enough, for (141) and (142):
(140) The speaker believes it is ridiculous to say that the bus is leaving.
(141) It is ridiculous to say that the bus is leaving.
(142) The bus is not leaving.
Or, to consider a final case, suppose there has been an argument about
when the bus would leave, with the speaker of (131) maintaining that it
will not leave for ten minutes and the hearer insisting that it will leave
immediately. When the bus moves off and the speaker says (131), the
assumption expressed by her utterance will be irrelevant to the hearer,
who is already aware that the bus is moving off. In these circumstances,
the description in (132) would achieve relevance not by providing the
hearer with indirect evidence for the assumption expressed, but by
providing him with evidence for such higher-level descriptions as
(143 )-(144):
(143) The speaker acknowledges that the bus is leaving.
(144) The speaker admits that she was wrong.
.t
:r
:r
,,.
There are thus a variety of ways in which a description such as (132) can
be relevant; some will have the effect of an ordinary assertion, others the
effect of a report of speech or thought, others the effect of an irony or
dissociation, others the effect of a speech-act classification and so on. A
speaker who wants to achieve some particular effect should give whatever
linguistic cues are needed to ensure that the interpretation consistent with
the principle of relevance is the one she intended to convey. Thus, when
an utterance is interpreted as an ordinary assertion, this is not a result of
the operation of some maxim of quality or convention of truthfulness,
but simply of an interaction between the form of the utterance, the hearer's
accessible assumptions and the principle of relevance. 3D
It is tempting to assume that there is an exactly parallel account of
imperatives to the one just proposed for declaratives, replacing the terms
'declarative form', 'saying that' and 'belief' by 'imperative form', 'telling
to' and 'desire', respectively. On this approach, an imperative utterance
such as (145) would be integrated into a description such as (146), which
250
Aspects
of verbal communication
could again be relevant in a variety of ways - for example, by providing

the hearer with evidence for assumption (147), on the basis of which he
might then form the desire to leave the room:
(145) Leave the room.
(146) The speaker is telling the hearer to leave the room.
(147) The speaker wants the hearer to leave the room.
In fact, the situation is slightly more complicated than these superficial
parallels might suggest. The problem is that there are many types of
imperative utterance which are used neithel' to express a desire of the
speaker's, nor to report on someone else's expression of desire. Compare
(148)-(149) with (150)-(151):
(148) Driver to traffic warden: Pretend you didn't see me.

(149) Keep my dog off his garden, he tells me. As if I could.
(150) (a) He: Could you tell me the way to the station?
(b) She: Turn right at the traffic lights and keep straight on.
(151) Recipe for mint sauce: Mix two tablespoons of mint leaves, two
teaspoons of sugar and half a tablespoon of hot water, add two
tablespoons of vinegar and leave to stand.
Whereas (148) is plausibly analysed as an expression by the driver of a
desire of her own, and (149) is plausibly analysed as a report by the speaker
of someone else's expression of desire, no parallel analysis is possible for
(150) and (151). There is no need for the hearer of (150b) to assume that the
speaker actually cares whether he turns right or not. There is no reason for
the reader of (151) to assume that the writer actually wants anyone who
sees the recipe to start making mint sauce. In these cases, the correlation
between imperative form and the propositional attitude of desire seems to
break down.
It might seem that at this point the speech-act framework comes into its
own. A speech-act theorist could ignore the possible links between
linguistic form and propositional attitudes such as belief and desire, and
simply note that the speech acts performed by imperative utterances fall
into two broad types: requestive, as in (148)-(149), and advisory, as in
(150)-(151). However, there is a problem with this proposal. Speech acts
in the advisory class - giving advice and making suggestions, for example
- surely do not have to be recognised as such in order to be performed. In
that case it is a mistake to offer an analysis on which the assignment of an
imperative utterance to the advisory or the requestive class of speech acts
would be fundamental to its comprehension.
We would like to suggest that the distinction between requestive and
advisory speech acts is itself reducible to something deeper. Intuitively, a
requestive speech act is one that represents a certain state of affairs as
Speech acts
251
desirable from the speaker's point of view, whereas an advisory speech act
is one that represents a certain state of affairs as desirable from the hearer's
point of view. What makes (148) above intuitively requestive is the fact
that the speaker is representing as desirable from her own point of view a
state of affairs in which the traffic warden pretends he did not see her;
what makes (150b) above intuitively advisory is the fact that the speaker is
representing as desirable from the hearer's point of view a state of affairs in
which the hearer turns right and keeps straight on. What is essential to the
comprehension of these utterances is not their assignment to the class of
advisory or requestive speech acts, but a recognition that the state of
affairs described is being represented as desirable from the speaker's point
of view in the first case, and the hearer's in the second.
If we are right, then the interpretation of imperative and declarative
utterances might proceed along broadly parallel lines. The hearer, on
recovering the propositional form P of an imperative utterance would integrate it into a description of the form The speaker is telling the hearer to P.
Telling the hearer to P might be analysed as communicating that the thought
that P interprets is entertained as a description of a desirable state of affairs.
Who entertains this thought in this way: the speaker or someone whose
thought the speaker is interpreting? From whose point of view is the state
of affairs described desirable? The hearer has to answer these questions
inferentially. As usual, the first interpretation consistent with the
principle of relevance will be selected, and a speaker who wants to be
correctly understood must make sure that the interpretation she intends to
convey is the first one consistent with the principle of relevance. We
believe that along these lines a satisfactory account of imperative
utterances might be constructed. On this account, the most basic, literal,
non-attributive imperatives would be the result of choosing path (d) in
figure 3 above and producing an utterance which was a literal interpretation of the speaker's thought. Metaphorical but non-attributive imperatives would be the result of choosing the same path but producing an
utterance which was a less than fully literal interpretation of the speaker's
thought. Attributive imperatives would be the result of choosing path (a).
Speech-act theorists tend to analyse interrogative utterances as a special
sub-type of directive speech act: specifically, as requests for information
(see Searle 1969: 69; Bach and Harnish 1979: 48). However, exam
questions such as (152), rhetorical questions such as (153), expository
questions such as (154), self-addressed questions such as (155) and indirect
questions such as (156) all present problems for this approach:
(152) What were the causes of the First World War?
(153) When did you say you were going to give up smoking?
(154) What are the main objections to this approach? First ...
252
(155) Why do the leaves of different trees go different colours in autumn?

(156) Peter doesn't know who his neighbour is.
When an examiner asks (152) above, it is not because she wants to know
the answer, but because she wants to evaluate the candidate's attempt at an
answer. A speaker who asks the rhetorical question (153) would not
normally be expecting any verbal response at all. A standard expository
device of many writers is to ask a question, such as (154), which they then
proceed to answer themselves. Many questjons, such as (155), are
produced in the absence of any audience, as pure intellectual speculations
or musings. Indirect questions such as (156) also resist speech -act analysis.
It is hard to see what request for information is being made, or even
alluded to, in (156): (156) could be true without it ever having occurred to
Peter to wonder, let alone ask, who his neighbour is. The standard
speech-act approach thus rules out any possibility of a unitary account of
direct and indirect questions.
We would like to suggest that an account of interrogative utterances can
be built around the notion of an interpretive use introduced in section 7.
Our hypothesis is that the hearer of an interrogative utterance recovers its
logical form and integrates it into a description of the form The speaker is
asking Wh-P, where Wh-P is an indirect question. Let us distinguish
between yes-no questions, which have not only a logical but also a fully
propositional form, and Wh-questions, which have a logical form but no
fully propositional form. Then we want to analyse asking Wh-P, where
Wh-P is a yes-no question and P is the propositional form of the
utterance, as communicating that the thought interpreted by P would be
relevant if true. We want to analyse asking Wh-P, where Wh-P is a
Wh-question and P is the less-than-propositional logical form of the
utterance, as communicating that there is some completion of the thought
interpreted by P into a fully propositional thought which would be
relevant if true. In other words, interrogative utterances are interpretations of answers that the speaker would regard as relevant if true.
Relevance, like desirability, is a two-place relation: what is relevant to
one person may not be relevant to another. Thus, in interpreting a
question, the hearer must always make some assumption about who the
speaker thinks its answer would be relevant to. Different assumptions
yield different types of question. For example, rhetorical questions such as
(153) above ('When did you say you were going to give up smoking?') are often
reminders, designed to prompt the retrieval of information the speaker
regards as relevant to the hearer. Similarly, expository questions such as
(154) above (,What are the main objections to this approach? First .. .'),
and more generally offers of information, are analysable as questions
whose answers the speaker regards as relevant to the hearer. Regular
requests for information, by contrast, are analysable as questions whose
Speech acts
253
answers the speaker regards as relevant to her, and believes, moreover,

that the hearer might be in a position to supply. In pure speculations such
as (155) above ('Why do the leaves of different trees go different colours in
autumn ?'), again the suggestion is that the answer would be relevant to the
speaker, but there is no manifest expectation that the hearer will be in a
position to supply it. In exam questions such as (152) above ('What were
the causes of the First World War?') the suggestion is that the answer will
be relevant to the speaker, not so much for its content as for the indirect
evidence it provides about the candidate's mastery of the subject. There
are thus a variety of ways in which the relevance of the description She is
asking Wh-P can be established, the non-attributive forms of which are
the result of choosing path (b) in figure 3 above.
Different questions can be relevant in different ways, some of which
have been sketched above. There is no need to analyse all questions as
requests for information, no need to set up special speech-act categories to
handle offers of information, rhetorical questions, expository questions
and so on. Questions can be successfully analysed without appeal to the
machinery of speech-act theory.
One advantage of this approach is that it suggests a way of explaining
the striking syntactic parallelisms between interrogative and exclamative
sentences (see Grimshaw 1979). In traditional speech-act terms, since
interrogatives are requests for infurmation and exclamatives are emphatic
assertions, it is hard to account for the consistent cross-linguistic
parallelisms between utterance types which have so little in common in
speech-act terms. Let us assume, however, that exclamatives, like
interrogatives, are specialised for interpretive rather than descriptive use,
and like non-attributive interrogatives are the result of choosing path (b)
in figure 3 above. Whereas a speaker who asks Wh-P (where Wh-P is an
indirect question) guarantees the relevance of some true completion of the
incomplete thought represented by P, a speaker who says that Wh-P
(where Wh-P is an indirect exclamation) guarantees the truth of some
relevant completion of the incomplete thought represented by P. On this
account, interrogatives and exclamatives have a lot in common.
Consider (157) and (158):
(157) Jane is so clever!
(158) How clever Jane is!
What we are suggesting is that the speaker of (157) or (158) guarantees the
truth of some relevant completion of the logical form she has expressed:
that is, of some assumption which would be relevant to the hearer and
which says how clever Jane is. Which assumption would that be? By the
general principles outlined above, it must be the first accessible assumption that is consistent with the principle of relevance. On this
254
analysis, the speaker of (157) and (158) is guaranteeing thatJane is cleverer

than the hearer would otherwise have expected. Thus, the intuition that
exclamatives are emphatic assertions and the striking parallelisms between
exclamative and interrogative form are simultaneously explained.
This very sketchy discussion of speech acts illustrates the general
relevance of the principle of relevance. The principle of relevance makes it
possible to derive rich and precise non-demonstrative inferences about the
communicator's informative intention. With the principle, all that is
required is that the properties of the ostensive stimulus should set the
inferential process on the right track; to do this they need not represent or
encode the communicator's informative intention in any great detail.
Thus, illocutionary-force indicators such as declarative or imperative
mood or interrogative word order merely have to make manifest a rather
abstract property of the speaker's informative intention: the direction in
which the relevance of the utterance is to be sought.
Postface
1 Introduction
In the nine years since Relevance was first published, the theory of
communication it proposes has been widely accepted, widely criticised
and widely misunderstood. The book has been translated into several
languages;! its implications for pragmatic theory have been explored in
a growing number of books and articles; it has inspired work in
neighbouring disciplines, including linguistics, literary studies, psychology and philosophy. In section 2 of this postface, we review briefly the
main developments that have taken place since the first edition was
published. 2
Many commentators, to whom we are very grateful, have raised a wide
variety of objections to the theory.3 We have had the opportunity to
answer most of them in a series of publications to which interested
readers are referred. 4 These criticisms have helped us correct some mistakes in the book; they have also made us aware of the difficulties in
comprehension and the many possibilities of misunderstanding it presents. Either because we are dense, or because we have had more time
than our commentators to think about these issues, we find that the most
serious problems with our theory are those we have discovered ourselves.
In section 3 of this Postface, we outline these problems, and propose
several significant changes both of formulation and of substance.
2 Developments
There is now a substantial body of work expounding and evaluating the
basic ideas of relevance theory. This includes a precis of Relevance,s two
textbooks and large sections of an encyclopaedia of pragmatics,6 expository articles designed for non-specialist audiences/ and several lengthy
256
Postface
critiques. s The implications of the theory have been explored in monographs and dissertations,9 there are edited collections including papers in
relevance theory,IO there is an e-mail network for exchange of ideas, and
a bibliography for classroom useY Several research projects have been
undertaken; informal workshops are held each year in London, and
more formal conferences and lecture series have been held around the
world. We will not attempt here a survey of this very diverse literature,
but merely point out some of the directions in which we feel that
particularly interesting and fruitful work is being done.
2.1 Explicit communication and the explicit-implicit distinction

Grice seems not to have noticed (or at least not to have developed the
idea) that his Co-operative Principle and maxims could help with other
aspects of pragmatic interpretation than the recovery of implicatures:
with disambiguation and reference assignment, for example, which he
saw as contributing not to what is implicated but to what is (explicitly)
said. In 'Logic and Conversation', he gives the impression that sentence
meaning and contextual factors are enough on their own to account for
disambiguation and reference assignment, and most Gricean pragmatists
simply followed him on this.12 This oversight had two important
consequences. First, Gricean pragmatists were slow to react to the
extensive psycho linguistic work being done on disambiguation and
reference assignmentY Second, they tended to take for granted that
pragmatic principles make no contribution to explicit content, and that
any aspect of utterance interpretation in which pragmatic principles play
a role is automatically an implicature. 14
In Relevance (chapter 4, section 2), we rejected this view of pragmatics
as de facto co-extensive with the study of implicatures. We introduced a
notion of explicature, parallel to Grice's notion of implicature, and a
definition of explicit communication, which we saw as 'richer, more
inferential and hence more worthy of pragmatic investigation than do
most pragmatists in the Gricean tradition'. A start was made on studying
disambiguation and reference assignment from a relevance-theoretic
perspective. We also questioned Grice's suggestion (1989: 25) that
disambiguation and reference assignment are the only context-dependent
processes involved in explicit communication, drawing attention to a
range of further inferential processes required to complete the interpretation of semantically incomplete expressions, narrow the interpretation
of vague expressions and, more generally, enrich the linguistically
encoded meaning to a point where the resulting overall interpretation
would be relevant enough.
The distinction between explicit and implicit communication, and the
Postface
257
role of pragmatic factors in eXp'licit communication, have been the focus

of much recent research. As noted in Relevance (chapter 4, section 3),
psycholinguists have provided valuable in sights into the actual processes
of disambiguation and reference assignment by investigating, for
example, how many candidate interpretations are activated, and at what
point one is selected and the others dismissed. However, they have been
less interested in what makes the selected interpretation pragmatically
acceptable, and on this point pragmatic theorists have a contribution to
make. Relevance theory claims that in disambiguation and reference
assignment, as in every other aspect of interpretation, the first interpretation consistent with the principle of relevance is the one the hearer
should choose. IS This is not the criterion suggested by most psycholinguists, who tend to talk in informal, Gricean terms. While pragmatic
theory can contribute to the development of an adequate criterion, it
also stands to gain from the fact that disambiguation and reference
assignment are more amenable to experimental testing than the recovery
of implicatures. Here, collaboration between pragmatists and psycholinguists should be of benefit to both.
Robyn Cars ton has studied the contribution of enrichment processes
to explicatures in a series of important papers;16 the role of inference in
explicit communication is now being actively explored both inside and
outside the relevance-theoretic framework. 17 Criteria have been proposed for distinguishing explicatures from implicatures, and a case has
been made for reanalysing some of Grice's best-known examples of
generalised implicatures (e.g. the temporal implicatures carried by
conjoined utterances, the quantity implicatures carried by numerals such
as 'two' and 'three') as pragmatically determined aspects of explicit
content. Much of this case rests on an intuitive distinction between
truth-conditional and non-truth-conditional content which is standard
throughout the speech-act and Gricean literature, but which could itself
be usefully reassessed.
The claim that pragmatic principles can contribute to explicit content
as well as implicatures has been seen as problematic by those who,
following Gazdar (1979), thought of the semantics-pragmatics distinction in a rather non-Gricean way. Gazdar imported into pragmatics a
picture common enough in formal semantics at the time, which conflated
linguistic semantics with truth-conditional semantics and defined pragmatics as 'meaning minus truth conditions'. On this account, pragmatic
processes should be 'post-semantic', and should not 'intrude' into the
truth-conditional domain.
Relevance theorists have consistently rejected this picture. 18 In Relevance (chapter 4, sections 1 and 7), following Fodor (1975), we
systematically distinguished between linguistic semantics (the semantics
258
Postface
of natural-language sentences) and truth-conditional semantics (the

semantics of conceptual representations). On this approach, the pragmatic processes that contribute to explicit truth-conditional content do
not 'intrude' into a unitary semantics: they act on the output of linguistic
semantics, enriching incomplete logical forms into fully propositional
forms which are in turn the bearers of truth conditions. The need for
some such distinction - which is not original to relevance theory - is
now widely accepted by those working both inside and outside the
relevance-theoretic framework.
2.2 Linguistic semantics
The implications of relevance theory for linguistic semantics, and in
particular for what is traditionally regarded as non-truth-conditional
linguistic meaning, have been a second major focus of research. In
previous frameworks, non-truth-conditional meaning was typically analysed in speech-act terms. Speech-act semanticists treated a range of nontruth-conditional expressions (mood indicators, discourse adverbials,
discourse particles, parentheticals) as indicators of illocutionary force.
Grice extended this account to a range of non-truth-conditional discourse connectives, which he treated as conventionally implicating the
performance of higher-order illocutionary acts. 19 Within the relevancetheoretic framework, this approach to non-truth-conditional meaning is
being reassessed. 20
Much of this reassessment was inspired by Diane Blakemore (1987),
who reanalysed Grice's discourse connectives using a distinction
between conceptual and procedural encoding; her account of discourse
connectives as encoding procedural constraints on implicatures has
provoked a flood of research. 21 A further impetus was provided by our
arguments against speech-act accounts of mood indicators in Wilson and
Sperber (1988a), and by our more general critique of speech-act theory
in Relevance (chapter 4, section 10).22
In Wilson and Sperber (1993), we argued that mood indicators and
discourse particles are best analysed in procedural rather than conceptual
terms. In the relevance-theoretic framework, both types of expression
contribute to explicatures rather than implicatures. We therefore generalised Blakemore's notion of constraints on implicatures, arguing that
procedural meaning can constrain any aspect of the inferential phase of
comprehension, whether explicit or implicit. We also questioned the
assumption that procedural meaning and non-truth-conditional meaning
invariably coincide: some expressions (e.g. discourse adverbials) which
are standardly treated as non-truth-conditional may be best seen as
encoding concepts; some truth-conditional expressions (e.g. pronouns)
Post/ace
259
may be best seen as encoding procedures. Relevance-theoretic alternatives

to speech-act accounts of mood indicators, discourse particles, discourse
adverbials and parentheticals sketched in that paper are now being actively
exploredY It may turn out that the conceptual-procedural distinction
will shed more light on linguistic semantics than the traditional distinction
between truth-conditional and non-truth-conditional meaning.
2.3 Interpretive dimensions
0/ language use
More fundamental than any of the distinctions discussed above is the

one drawn in Relevance (chapter 4, sections 7-9), between descriptive
and interpretive dimensions of language use. We claimed that, on the
most basic level, every utterance is a more or less faithful interpretation
of a thought the speaker wants to communicate. An utterance is
descriptively used when the thought interpreted is itself entertained as a
true description of a state of affairs; it is interpretively used when the
thought interpreted is entertained as an interpretation of some further
thought: say, an attributed or a relevant thought. In the light of this
distinction, traditional pragmatic categories, e.g. tropes and speech acts,
must be radically rethought: for example, metaphor falls together with
descriptive uses of language, while irony, interrogatives and exclamatives
fall together as varieties of interpretive use.
Our approach to metaphor and irony, developed in a series of later
papers, has been extensively discussed. 24 Perhaps surprisingly, most
reactions have come not from Gricean pragmatists, whose analyses we
severely criticised, but from psychologists, non-Gricean pragmatists and
literary theorists. The range of data now being considered, and the range
of explanations on offer, are much richer than those discussed in the
rather limited Gricean literature.
The interpretive dimension of language use is not restricted to irony.
Translation has been reanalysed from this perspective in a series of
interesting works by Ernst-August Gutt. 25 The notion of interpretive
use has also shed light on a range of traditional linguistic topics such as
interrogatives, exclamatives, echo questions, pseudo-imperatives, hearsay particles and metalinguistic negation, most of which have resisted
analysis in purely descriptive terms. 26 There is much more to be done in
this area, from both descriptive and theoretical points of view. However,
the reorganisation proposed in Relevance seems to be bearing fruit.
2.4 Wider domains

A start has been made on investigating the implications of relevance
theory in wider domains. In literary studies, the suggestions made by
260
Postface
Paul Kiparsky (1987) have been actively pursuedY Humour, politeness,

advertising, argumentation, political language and language in education
have all been investigated from a relevance-theoretic perspective. 28 Ruth
Kempson has applied the assumptions of relevance theory to the
investigation of generative grammar and issues of linguistic modularity.29
Foster-Cohen (1994) and Watson (1995) have looked at language
development; the broader implications of relevance theory for language
acquisition are assessed in Smith (1989), Smith and Tsimpli (1995);
theoretical considerations bearing on both evolution and development
are discussed in Sperber (1994a).
:
In psychology, interesting results are being obtained in several
domains. Frith (1989) and Happe (1991, 1992, 1993) have applied
relevance theory to the analysis of autism. Politzer (1993) has reanalysed
several major experimental paradigms in the psychology of reasoning,
and shown how considerations of relevance affect the performance of
subjects in ways that can explain some of the most striking experimental
results. Sperber, Cara and Girotto (forthcoming) have reanalysed the
literature on Wason's famous Selection Task, where subjects are asked
to select evidence potentially relevant to evaluating the truth of a
conditional statement. Sperber et al. suggest that the performance of
subjects can be explained on the basis of intuitions of relevance
developed in the process of comprehending the task. Their analysis
yields precise and novel predictions involving the manipulation of effect
and effort, which have been experimentally confirmed.
3 Revisions
3.1 Not one but two Principles of Relevance

In Relevance, we make two fundamental claims, one about cognition,
the other about communication:
(1) Human cognition tends to be geared to the maximisation of
relevance.
(2) Every act of ostensive communication communicates a presumption
of its own optimal relevance.
Claim (2) is what we called the Principle of Relevance. However, many
readers, even careful ones, have used the term 'Principle of Relevance' to
refer to claim (1). This is a straight misreading, but an understandable one.
Claim (1) is more fundamental and general than claim (2), and at least as
worthy to be called a principle. We originally called claim (2) a principle
to contrast it with other pragmatic 'principles' proposed in the literature:
in particular Grice's Co-operative Principle. We failed to foresee that
Postface
261
when our book was read and interpreted - as we wanted - in the context
of wider cognitive concerns, this use of the term 'principle' would seem
rather arbitrary, cause unnecessary effort, and hence (as we should have
predicted on relevance-theoretic grounds) lead to misinterpretation.
We have decided to remedy the situation by talking in future of two
Principles of Relevance: the First (or Cognitive) Principle is given in (1),
and the Second (or Communicative) Principle is given in (2). Throughout this book, the term 'Principle of Relevance' refers to the Second,
Communicative Principle. The change is, of course, expository and not
substantive, but it is worth spelling out what we hope to highlight by
this reformulation.
3.2 The First Principle of Relevance
The First Principle of Relevance is less subtle than the Second Principle,
but it is still controversial and in need of justification. As stated, it is
also too vague, and in need of elaboration.
Relevance is not a commodity; it is a property. What is it a property
of? By our definition, it is a property of inputs to cognitive processes. It
can be a property of stimuli, for example, which are inputs to perceptual
processes, or of assumptions, which are inputs to inferential processes.
Stimuli, and more generally phenomena, are found in the environment
external to the organism; assumptions, which are the output of cognitive
processes of perception, recall, imagination or inference, are internal to
the organism. When we claim that human cognition tends to be geared
to the maximisation of relevance, we mean that cognitive resources tend
to be allocated to the processing of the most relevant inputs available,
whether from internal or external sources. In other words, human
cognition tends to be geared to the maximisation of the cumulative
relevance of the inputs it processes. It does this not by pursuing a longterm policy based on computation of the cumulative relevance achieved
over time, but by local arbitrations, aimed at incremental gains, between
simultaneously available inputs competing for immediately available
resources.
Why assume that human cognition tends to be geared to the maximisation of relevance? The answer comes in two stages, one to do with the
design of biological mechanisms in general, the other with efficiency 'in
cognitive mechanisms.
We start from the assumption that cognition is a biological function,
and that cognitive mechanisms are, in general, adaptations. As such,
they are the result of a process of Darwinian natural selection (although
other evolutionary forces may have helped to shape them). We assume,
then, that cognitive mechanisms have evolved in small incremental steps,
262
Postface
mostly consisting in the selection of a variant that performed better at

the time than other variants that were around. There are many ways in
which one variant of a biological mechanism can perform better than
others. There may be a qualitative difference in the type of benefits that
different variants produce; or the difference may be quantitative, as
when the same benefit can be achieved to a greater degree, or at a lower
energy cost.
Whereas selection pressures for qualitative improvements vary perpetually with changes in the genotype and the environment, selection
pressures for quantitative improvements are' a relatively stable factor.
Ceteris paribus, greater benefits or lower costs are always a good thing.
In principle, there are many equally satisfactory ways of balancing costs
and benefits: many ways, that is, of being efficient (although few, if any,
may be genuine alternatives at a given point in the evolution of an
adaptation). Hence, it is not possible to predict what exact balance of
costs and benefits should be achieved in a given biological mechanism as
a result of the pressure towards greater efficiency. What we can expect
is that, in general, an enduring biological mechanism with a stable
function will have evolved towards a better cost-benefit balance, i.e.
towards greater efficiency.
For example, we can expect that the structure, placement and mode
of operation of a muscle will tend to minimise the energy costs of
performing the bodily movement it is its function to produce. Similarly,
we can expect to find a tendency towards maximal efficiency in the
design of cognitive mechanisms.
We assume, too, that human cognition is the joint product of many
specialised mechanisms (see Barkow, Cosmides and Tooby 1992;
Hirschfelf and Gelman 1994). Each cognitive mechanism contributes its
qualitatively different benefits, in the form of cognitive effects. For each,
there has been pressure towards cost-benefit optimisation.
All these cognitive mechanisms taken together constitute the cognitive
system. The efficiency of the cognitive system as a whole depends on
how its various sub-mechanisms are articulated with one another, and
how the resources of the system are shared among them. Articulation
and allocation of resources must be such as to maximise the likelihood
that the most relevant available information will be processed in the
most relevant way.
What the First Principle of Relevance says is that human cognition
tends to be organised so as to maximise relevance. There may be many
shortcomings, many cognitive sub-mechanisms that fail to deliver
enough effect for the effort they require, many occasions when the
system's resources are poorly allocated. The First Principle does not
rule these out. Still, for it to be of any use, the tendency towards
Pastface
263
maximi~ation o~ relevance must be strong enough overall to help guide
human mteractIOn. After all, The Second, Communicative Principle of

~elevance is gr~)Unde~ i~ the First.Principle, and in the further assumptIOn that the FIrst PnnCIple does mdeed make the cognitive behaviour
of another human predictable enough to guide communication.
3.2.1 The First Principle of Relevance and truth Our definition of the
relevance of an assumption in a context takes no account of the objective
truth or falsity of the assumption itself, or of the conclusions that may
be derived from it in the context. Thus, a false assumption that
contextually implies many false conclusions, or a true assumption that
combines with a false contextual premise to imply many false conclusions, is, by our definition, as relevant as a true assumption that
implies many true conclusions. On the other hand, our rationale for
introducing this notion of relevance has to do with considerations of
cognitive efficiency, and the notion of cognitive efficiency cannot be
divorced from that of truth. The function of a cognitive system is to
deliver knowledge, not false beliefs. Does this mean there is something
missing from our definition of relevance? Definitely, and it is in need of
revision. Note, though, that for most of our purposes our incomplete
definition is good enough.
When we use the notion of relevance to help describe how a cognitive
system allocates its resources, there is no harm in leaving objective truth
or falsity out of account. The system has no other way of distinguishing
true from false assumptions than via its own inputs and internal
processes. Basically, if an assumption is caused by the environment in
the appropriate way (e.g. through perception), the system accepts it; if
an assumption is inferentially derived by the system~s own computational mechanisms from accepted premises, it again accepts it. When the
system is a reflective one, e.g. a human being, it may be aware that it
wants real knowledge and not false beliefs; it may be aware of the risk
of accepting false assumptions; it may develop some procedures to
double-check the outcome of other procedures; but all it can do in the
end is trust the sum of its own procedures to deliver knowledge. So the
system will take the output of its own mechanisms as cognitively
warranted, and will assess relevance in terms of all contextual effects
achieved, even though, unbeknownst to it, some of its conclusions may
turn out to be false. From this solipsistic point of view (in the sense of
Fodor 1980), truth can safely be ignored.
However, this is not the only point of view that needs to be taken
into account. A reflective cognitive system may be aware that some of
its beliefs are likely to be false, even if it cannot tell which, and it may
regard information leading to false beliefs as worse than irrelevant.
264
Postface
Similarly, a reflective cognitive system that communicates with other

systems may regard only true information as relevant to them. Take a
speaker who wants her audience to think she is married, when in fact
she is not. She lies:
(3) I am married.
Does she believe that what she says is relevant to the hearer, or only
that it may seem relevant to him, since it would have been relevant if
true? We suggest the latter.
Relevant information is information worthChaving. False information
is generally not worth having; it detracts from cognitive efficiency. How
should we incorporate this epistemic feature into our definitions? There
are two possibilities: we might say that inputs to cognitive processes are
relevant only if they meet some specific epistemic condition; or we
might say that inputs are relevant only if the outputs of their cognitive
processing meet some specific epistemic condition.
The most obvious, and apparently simplest, solution is to make truth
of the input a necessary condition of relevance. There are three problems
with this choice. First, we want to attribute relevance not just to
assumptions but also to phenomena, and in particular to ostensive
stimuli. These are inputs to cognitive processes, but they are not the sort
of things that can be true or false. Utterances, of course, are said to be
true or false, and they are a kind of ostensive stimulus; but when we say
that an utterance is true, we really mean that its interpretation is true,
and this is the output of a cognitive process of comprehension.
Second, truth of the conclusions seems more crucial to relevance than
truth of the premises. Consider the following scenarios:
(4) Peter is a jealous husband. He overhears Mary say on the phone to
someone, 'See you tomorrow at the usual place.' Peter guesses
rightly that she is speaking to a man, and infers, quite wrongly, that
she has a lover and does not love him any more.
(5) Peter is a jealous husband. He overhears Mary say on the phone to
someone, 'See you tomorrow at the usual place.' Peter guesses
wrongly that she is talking to a man, and infers, rightly as it happens,
that she has a lover and does not love him any more. (Mary's lover
is a woman.)
In (4), Peter's assumption that Mary was talking to a man was true, and
led to rich contextual effects. However, these effects were false beliefs.
Was Peter's assumption relevant? We would rather say that it seemed
relevant, but in fact was not. In (5), by contrast, Peter's assumption that
Mary was talking to a man was false, but it led to many true beliefs, so
that here we would be willing to say that it was genuinely relevant
Post/ace
265
(though perhaps not as relevant as it seemed, since it also led to some

false beliefs).
Take the more general case of fiction. When you hear a parable, or
read War and Peace, you may gain insight, through some form of
analogical thinking, into yourself, your life, and the world as they are.
If only true inputs were relevant, we would have to say such fictions
were irrelevant. If truth of the output is what matters, then fictions can
be relevant after all.
So let us explore the second way of amending our definition of
relevance: by treating an input as relevant only if the output of its
cognitive processing meets some specific condition. The basic idea is
that for an input to be relevant, its processing must lead to cognitive
gains. Now recall our strategy in the book. We first defined relevance in
a context, and then relevance to an individual. Our definition of
relevance in a context can be left unchanged. A context, even coupled
with an inference engine, is not yet a cognitive system; it does not have
a cognitive function, and does not stand to gain from true representations or lose by false ones. Relevance in a context is a formal property,
interesting as such (with possible applications in Artificial Intelligence,
for instance), and is best left as it is.
Things change when we move from relevance in a context to relevance
to an individual (or more generally, to any cognitive system). Contextual
effects in an individual are cognitive effects (a phrase we have used in
articles written after 1986). They are changes in the individual's beliefs.
An individual does stand to gain or lose by the truth or falsity of his
beliefs, and he does have cognitive goals. An individual, were he to reflect
on it, would not be interested in contextual effects per se, but only in so
far as they contribute to his cognitive goals. This is easily built into our
definition of relevance to an individual. Let us first define a cognitive
effect as a contextual effect occurring in a cognitive system (e.g. an
individual), and a positive cognitive effect as a cognitive effect that
contributes positively to the fulfilment of cognitive functions or goals.
Then we replace definitions (42) and (43) of chapter 3 with (6) and (7):
(6) Relevance to an individual (classificatory)

An assumption is relevant to an individual at a given time if and
only if it has some positive cognitive effect in one or more of the
contexts accessible to him at that time.
(7) Relevance to an individual (comparative)
Extent condition 1: An assumption is relevant to an individual to the
extent that the positive cognitive effects achieved when it is optimally
Extent condition 2: An assumption is relevant to an individual to the
266
Postface
extent that the effort required to achieve these positive cognitive
effects is small.
Definitions (58) and (59) of the relevance of a phenomenon to an

individual should also be modified accordingly.
These changes in the definition of relevance might seem to raise two
questions. First, isn't the notion of a positive cognitive effect far too
vague ? Well, we could have been more specific and defined a positive
cognitive effect as an epistemic improvement, i.e. an increase in knowledge. All the effects we are actually considering in this book are of this
relatively well-defined epistemic kind. However, we want to leave open
the possibility of taking into account, in the full picture, other possible
contributions to cognitive functioning, involving, for instance, the
reorganisation of existing knowledge, or the elaboration of rational
desires. And, yes, the resulting definition of a positive cognitive effect is
vague, but that is a problem not for relevance theory, but for cog~itive
psychology in general.
The second question that this redefinition of relevance to an individual
might seem to raise is this. Doesn't the First Principle of Relevance then
become vacuous? If human cognition tends to be geared to the maximisation of relevance, and if relevance is itself defined in terms of positive
cognitive effects, aren't we just saying that human cognition tends to be
geared towards the production of positive cognitive effects; which,
surely, is a truism, and a vague one at that?
In fact, the First Principle is far from a truism. It makes two empirical
claims: neither is self-evident, and the second is original to relevance
theory. The First Principle might be false: human cognition might
achieve a balance of positive versus negative cognitive effects just good
enough to avoid being selected out. In fact, human cognition, being an
evolved and adapted system, reflects in fine-grained aspects of its design
repeated past pressures towards optimisation. Moreover, we claim that
there is one general and essential way in which human cognition exhibits
good design, and that is by tending to allocate its resources to the
processing of available inputs in such a way as to maximise the expected
cognitive effects. That said, we ourselves have stressed that what we
now call the First Principle of Relevance is indeed vague and general,
and that what makes it worth stating are some of its precise and nontrivial consequences: in particular, the Second Principle of Relevance.
3.3 Revising the presumption of relevance
The (Second) Principle of Relevance states that every act of ostensive communication communicates a presumption of its own optimal
Postface
267
relevance. The presumption of relevance itself was spelled out as

follows:
make manifest to the addressee is relevant enough to make it
worth the addressee's while to process the ostensive stimulus;
We believe that this formulation should be substantively modified. The

modifications will make the presumption of relevance simpler, and we
will argue that they not only preserve the predictive power of the earlier
version, but significantly increase it.
There are two reasons for crediting a communicator with the intention
to convey a presumption of relevance; these are reflected in the two
clauses of the presumption. First, the communicator must intend her
ostensive stimulus to appear relevant enough to the addressee to be
worth his attention. Otherwise, he might not pay it enough attention,
and communication would fail. This sets a lower limit on the level of
relevance the communicator intends the addressee to expect. A version
of this idea is built into clause (a) of the presumption of relevance above.
In this version, the level of effort needed to reconstruct the intended
interpretation is treated as given, and the presumption is that the effect
will be high enough for the overall relevance of the stimulus to be at or
above the lower limit (below which the stimulus would not be worth
processing). Clause (a) says, in essence, that the level of effect is at least
sufficient.
Now suppose we treat the level of effect rather than effort as given.
Then by the same reasoning - based on the fact that the communicator
must intend her ostensive stimulus to appear relevant enough - the
addressee can have legitimate expectations about the level of effort
needed to achieve this effect. This level of effort must be low enough for
the overall relevance of the stimulus to be at or above the lower limit.
Since there is no principled asymmetry here between effect and effort,
clause (a) of the presumption of relevance can be made both simpler and
more general, as follows:
(9) The ostensive stimulus is relevant enough for it to be worth the
addressee's effort to process it.
Is it ever legitimate for the addressee to expect - and the communicator to intend him to expect - a level of relevance that is not merely at
but well above the lower limit? Grice and most of his followers suggest
that it is. They assume that speaker and hearer must have a common
268
Post/ace
goal that goes beyond merely understanding and being understood, and
are expected to provide whatever information would best further this
common goal. What is to be expected is not just relevance enough, but
maximal relevance to achieving the common goal. 30
We have expressed disagreement with this view. It may be true that in
most verbal exchanges the participants share a purpose that goes beyond
merely understanding one another, but it need not always be the case.
Conflictual or non-reciprocal communication, for example, involve no
such purpose. It is also true that understanding is made easier by the
presence of a common goal. We can accountcfor by this pointing out
that a common goal creates a number of mutually manifest contextual
assumptions on which the interlocutors can draw. The existence of a
common conversational goal need not be built into pragmatic principles.
We still believe this is correct.
However, we ourselves have stressed that interlocutors always share
at least one common goal, that of understanding and being understood.
It is in the communicator's manifest interest both to do her best and to
appear to be doing her best to achieve this common goal. This provides
a second reason for crediting her with the intention to convey a
presumption of relevance, and is reflected in clause (b) of the presumption as stated above. In its current version, however, clause (b) is wholly
about effort. The intended effect is treated as given, and clause (b) says
that the stimulus used to achieve this effect is the one that requires least
effort from the addressee.
The presumption of minimal effort expressed by (b) is at best too
vague and at worst too strong. A communicator may well be willing to
try to minimise the addressee's effort, since this will make him more
likely to attend to her ostensive stimulus and succeed in understanding
it. Still, for all sorts of reasons, the particular stimulus she produces may
not be the one that would absolutely minimise the addressee's effort. In
the first place, there is the communicator's own effort to consider. As
speakers, we are prepared to make only so much effort in formulating
our thoughts, and as hearers, we know better than to expect flawlessly
crafted utterances. Then there may be rules of etiquette or standards of
ideological correctness that rule out the utterance that would be easiest
to process (which would also be likely to convey unwanted weak
implicatures). As speakers, we avoid what we see as objectionable
formulations, and as hearers, we expect such restraint.
Clause (b) of the presumption of optimal relevance should in any case
have allowed for the speaker's right to be lazy or prudish, i.e. to have
her own preferences and take them into account. 31 In later publications
or oral presentations, we amended this effort clause to say that no
unjustified or gratuitous effort was to be demanded. In other words,
Post/ace
269
from a range of possible stimuli which were equally capable of communicating the intended interpretation and equally acceptable to the
communicator (given both her desire to minimise her own effort and
her own moral, prudential, or aesthetic preferences), the communicator
should prefer, and appear to prefer, the stimulus that would minimise
the addressee's effort.
However, this line of reasoning, which was based on considerations
of effort, applies equally to the effect side. Suppose that, from the
communicator's point of view, her goals would be equally well served
by a number of utterances (or other stimuli), all of which would cause
the intended contextual effects, but some of which would cause further
contextual effects, and be (or seem) more relevant to the addressee as a
result. Which should she choose? She should choose the utterance that
would be (or seem) most relevant to the addressee, for just the reasons
given above in discussing the minimisation of effort.
Here is an illustration. Mary wants to make it quite manifest to Peter
that she will be out from 4 o'clock to 6 o'clock. She might inform him
of this by saying any of (lOa-c):
(10) (a) I'll be out from 4 to 6.
(b) I'll be out at the Jones's from 4 to 6.
(c) I'll be out at the J ones's from 4 to 6 to discuss the next meeting.
Suppose she assumes that any of these utterances would be relevant
enough to Peter. Suppose it doesn't matter to her whether she tells him
where she is going and why. Suppose the amount of effort needed to
produce any of these utterances makes no difference to her. Then it
would be rational enough to utter any of (lOa-c), since each would
achieve her goal at an equally acceptable cost to her. However, it would
be most rational to produce the utterance most relevant to Peter, since
this would make it most likely that he would attend to her communication, remember it, and so on: in other words, it would maximise the
manifestness to Peter of the information that Mary wants him to have.
Since (lOc) would demand more effort from Peter than (lOb), and (lOb)
than (lOa), Mary should choose one of these longer utterances if and
only if the extra information conveyed yields enough effect to make it
more relevant to Peter. If he doesn't care where she is going, she should
choose (lOa). If he cares where she is going, but not why, she should
choose (lOb). If he cares both where and why, she should choose (lOc).
These choices are rational even if Mary doesn't particularly want to be
helpful to Peter by telling him what he may want to know. They are
rational as ways of maximising the chances that she will succeed in
making manifest to him the one thing she does want to make manifest:
that she will be away from 4 o'clock to 6 o'clock.
270
Postface
We can thus make the following generalisation. Take a set of stimuli

that meet the following conditions: any of them would be likely to
communicate what the communicator wants to communicate; she
is capable of producing any of them; and she has no preferences
among them, apart from wanting to choose the one that will be most
effective in achieving her communicative goal. These stimuli may differ
in terms of the effort demanded of the addressee, the effects achieved,
or both effect and effort. The communicator should choose the stimulus
that appears most relevant to the addressee, since this will make
her communication most likely to succeed~ For the same reason, she
should appear to be choosing the stimulus that is most relevant to the
addressee. In normal conditions, appearance and reality are likely to
coincide.
The communicator's choice of ostensive stimuli is limited not only by
her preferences but by her abilities. On the effort side, there may be
stimuli that would be easier for the hearer to process, but that the
communicator is unable to think of at the time, as when the best
formulation of some thought just fails to come to mind. On the effect
side, the limits on the communicator's abilities are even more significant.
There may always be information that the hearer would find more
relevant than anything the communicator has to offer. She cannot be
more relevant than her own knowledge permits. If she decides to
communicate in bad faith, and tries to make manifest assumptions that
she does not believe, she would still want the addressee to think that
what she is trying to communicate is warranted by what she knows.
Again, there is no principled asymmetry between effect and effort.
The presumption is that, of all the stimuli that are available to her and
acceptable as a means of achieving her particular communicative goal,
the communicator will choose one that is as relevant as possible to the
addressee. The second clause of the presumption of relevance can be
made both simpler and more general, as follows:
(11) The ostensive stimulus is the most relevant one compatible with
the communicator's abilities and preferences.

We now have a fully revised presumption of optimal relevance:
(12) Presumption of optimal relevance (revised)
(a) The ostensive stimulus is relevant enough for it to be worth the
addressee's effort to process it.
(b) The ostensive stimulus is the most relevant one compatible
with the communicator's abilities and preferences.
This says that the addressee is entitled to expect a level of relevance high
enough to warrant his attending to the stimulus, and which is, moreover,
Postface
271
the highest level of relevance that the communicator was capable of

achieving given her means and goals.
3.3.1 The Second Principle of Relevance: that the presumption of optimal

relevance is ostensively communicated It would be a mistake to read
the presumption of optimal relevance, in either the early or the revised
version, as describing a goal that rational communicators should achieve.
Unlike Grice's maxims, neither the principle nor the presumption of
relevance is presented as a goal to be pursued or a rule to be followed
by the communicator. The (Second) Principle of Relevance is a descriptive (as opposed to normative) claim about the content of a given act of
ostensive communication. It claims that part of that content is a
presumption that this very act of communication is relevant to the
addressee.
The addressee's aim in interpreting an utterance is to identify the
communicator's informative intention. As with any attribution of an
intention to an agent, this is done by observing the means she chooses
and assuming that these are appropriate to her goals, given her beliefs.
We claim that a presumption of optimal relevance is communicated by
any act of ostensive communication. Given our definition of ostensive
communication, for this to be true it must be mutually manifest to
communicator and addressee that the communicator has the informative
intention of making the presumption of relevance mutually manifest.
We will now show that this is so.
A rational communicator must intend the stimulus she uses to appear
relevant enough to the addressee to attract his attention and make him
willing to spend the effort needed for comprehension. How relevant is
that? There is a limit below which the addressee will be unlikely to
attend to the stimulus at all; clearly, the communicator must intend the
addressee to expect a level of relevance at least as high as this. Moreover,
it is to the communicator's advantage that the addressee should expect a
level of relevance well above this lower limit, so that he will be willing
to invest the effort needed for comprehension. However, just as the
addressee is guided in interpreting the utterance by the assumption that
the communicator is rational, so the communicator's intentions are
constrained by the assumption that the addressee is rational. A rational
addressee will not expect more relevance than the communicator is
willing and able to achieve. There is no point in expecting the communicator to give information she doesn't have, or to produce stimuli she
is unable to think of at the time. Nor can she be expected to go against
her own preferences. So a rational communicator intends her stimulus
to appear as relevant as is compatible with her abilities and preferences.
In other words, it is necessary for the first clause of the presumption
272
Postface
of relevance to be manifest to the addressee, and it is advantageous for

the second clause to be manifest too. A rational communicator should
therefore want both clauses of the presumption of relevance to be
manifest. We claim that this is not some hidden fact about the psychology of communicators, but is manifest to any competent communicator
or addressee. Thus, when a communicator makes it mutually manifest
to herself and her addressee that she is trying to communicate by means
of a given stimulus, she thereby makes it mutually manifest that she
intends a presumption of relevance to be manifest. Given our definition
of ostensive communication, this amounts to Jsaying that a presumption
of relevance is communicated.
3.3.2 Some consequences of the revised presumption of relevance All
the analyses we have given in this book and elsewhere on the basis of
the old presumption of optimal relevance go through as before. It is still
true that the rational way to go about interpreting an utterance, or any
other ostensive stimulus, is to follow a path of least effort and stop at
the first interpretation that satisfies one's expectation of relevance.
However, in the old version, the expected level of relevance was
systematically at the lower limit. This did not mean that an utterance
could never be more than just relevant enough to be worth the hearer's
attention. What it did mean is that in order to achieve a higher level of
relevance, the speaker had to formulate her utterance so that the first
interpretation that was relevant enough to be worth the hearer's
attention would actually be more than relevant enough.
To illustrate, suppose that Mary says to Peter:
(13) You remember I bought that lottery ticket? Well, guess what? I
won 10,000!
Mary's statement, taken literally, may well be not only relevant enough
to be worth Peter's attention, but much more relevant than he would
have expected, given the unrevised presumption of relevance. Still, if this
is the first accessible interpretation that is relevant enough (and unless it
conflicts with other of his contextual assumptions), he will accept it as
the one intended. This, at least, is what an analysis based on the
unrevised presumption of relevance would (correctly) predict.
Compare this with the case where Mary says to Peter:
(14) You remember I bought that lottery ticket? Well, guess what? I
won a prize!
Here, the first accessible interpretation that is relevant enough will
probably represent Mary's prize as just big enough to be worth talking
about. If just knowing that she won a prize is relevant enough, then the
Postface
273
value of the prize may not be seen as relevant at all. Here again, an
analysis based on the unrevised presumption of relevance is adequate.
The revised presumption of relevance yields the same analysis of these
and similar examples. In interpreting (13), Peter assumes that Mary had
the ability - in this case the knowledge - to say something more than
minimally relevant (namely that she had won 10,000), and that she gave
this information in the absence of contrary preferences. In interpreting
(14), let us assume that Peter accepts clause (b) of the presumption of
relevance and expects Mary's utterance to be the most relevant one
compatible with her abilities and preferences. Still, he has no reason to
think she has a more relevant piece of information that she is reluctant
to share with him; so he will assume that the prize is merely big enough
to be worth mentioning. Quite often, the lower limit mentioned in
clause (a) of the (revised) presumption of relevance will coincide with
the higher limit mentioned in clause (b). The speaker has something just
relevant enough to be worth saying, and says it.
In some cases, though, the revised presumption yields different, and
better, analyses. Here we will consider two. The first is adapted from
Grice (1989: 32). Peter and Mary are planning a holiday in France. Peter
has just said that it would be nice to visit their old acquaintance Gerard
if it would not take them too far out of their way. The dialogue
continues:
(15) (a) Peter: Where does Gerard live?
(b) Mary: Somewhere in the South of France.
As Grice notes, Mary's answer implicates (16):
(16) Mary does not know where in the South of France Gerard lives
This implicature is easily explained in terms of Grice's maxims. Mary's
answer is less informative than the first maxim of Quantity ('Make your
contribution as informative as is required') would suggest. 'This
infringement [... ] can be explained only by the supposition that [Mary]
is aware that to be more informative would be to say something that
infringed the second maxim of Quality, "Don't say what you lack
evidence for'" (Grice 1989: 32-33).
In the unrevised version of relevance theory, we would have to explain
this implicature by noting that, in the situation described, it would
generally be mutually manifest that Mary is expected and willing to
co-operate in planning the holiday in France. From this assumption, together with the fact that her reply is not relevant enough to
answer Peter's question, it can be inferred that she does not know
exactly where Gerard lives. Then not only is (16) manifest but, given
274
Postface
Mary's co-operativeness, it is mutually manifest that she should want

(16) to be manifest. Hence (16) is a proper implicature.
This analysis acknowledges the presence, in this particular situation,
of the kind of co-operativeness which Grice regards as there in principle
in every conversation. We have argued that Gricean co-operativeness is
neither always at work, nor always presumed to be at work. In
circumstances where the speaker is not expected to be co-operative,
implicatures of the type in (16) do not go through.
Suppose, for example, that it is mutually manifest that Mary is dead
against visiting Gerard. Then her answer would"not carry the implicature
in (16). She may have no more precise information about Gerard's
whereabouts, or she may have it but be reluctant to give it, and there is
no telling which. Here, a strict Gricean would have to say that Mary is
at least partially 'opting out' of the Co-operative Principle and the first
maxim of Quantity. Just as we would have had to explain the Gricean
implicature in (16) by adding the contextual assumption that the speaker
is co-operative, so a Gricean would have to explain the absence of the
implicature by adding the assumption that the speaker is unco-operative.
Notice, now, that the same dialogue could carry a different implicature. Suppose it is mutually manifest that Mary knows where Gerard
lives. Then her answer in (lSb) would implicate not (16) but (17):
(17) Mary is reluctant to say exactly where Gerard lives.
This raises a problem for the Gricean, since it violates both the Cooperative Principle and the first maxim of Quantity, and implicatures
are supposed to arise only on the assumption that the Co-operative
Principle is in force. With the unrevised version of the presumption of
relevance, this example would have raised a problem for us too. Let us
suppose that the information that Gerard lives in the South of France is
relevant enough to be worth Peter's attention (even though it is less
relevant than he would wish). Then, on our unrevised account, Peter
should stop short of constructing the implicature in (17).
With the revised presumption of relevance, we can explain both
standard Gricean implicatures such as (16), and non-Gricean implicatures such as (17), which are caused by, and express, a refusal to cooperate. In (15), if it is mutually manifest that Mary would like to be
more specific about where Gerard lives, then her response, together
with clause (b) of the revised presumption of relevance, will imply that
she is unable to be more specific. If it is mutually manifest that this
implication increases the relevance of her utterance, then it will be not
just implied but implicated. On the other hand, if it is mutually manifest
that Mary could have been more specific, then her response, together
with clause (b) of the presumption of relevance, will imply that she is
Postface
275
unwilling to be more specific. Again, if it is mutually manifest that this

implication increases the relevance of her utterance, it will be implicated.
Note that here we are making a subtle and non-obvious claim. We are
claiming that if it is mutually manifest to communicator and audience
that an assumption contextually implied by an utterance increases its
overall relevance, then it is (in general) mutually manifest that the
communicator intended this implication to be manifest. In other words,
this assumption is communicated (as an implicature). This follows from
clause (b) of the revised presumption of relevance, which states that the
utterance is the most relevant one compatible with the communicator's
abilities and preferences. If a mutually manifest implication of the
utterance contributes to overall relevance, and thus helps to confirm the
presumption of optimal relevance, the inference that the communicator
intended it to play this role is sound. It is obvious that the communicator
is able to implicate this assumption. There is evidence that she is willing
to implicate it, since she has willingly chosen a form of utterance that
manifestly carries this implication, which helps to confirm the presumption of relevance that she herself has communicated.
The claim that manifestly relevant implications can be treated as
implicatures has one striking consequence. Sometimes, the addressee
may justifiably attribute to the communicator an implicature that she
never in fact intended to communicate. Sound though it may be, the
inference from the mutually manifest fact that an implication is relevant
to the conclusion that it is implicated (i.e. intentionally made manifest)
is a non-demonstrative one, and it may on occasion be false. Consider a
slightly different version of dialogue (15) above. It is mutually manifest
to Mary and Peter that Mary is willing to give him all the relevant
information she has:
(18) (a) Peter: You said you were in touch with Gerard. Where does
he live?
(b) Mary: Somewhere in the South of France, I don't know exactly
where.
In (18b), Mary says that she doesn't know exactly where Gerard lives.
As it stands, this utterance, made without further explanation, contextually implies that she misinformed Peter when she claimed to be in
touch with Gerard. She might not have intended to make this implication
manifest, and a fortiori she might not have wanted to implicate it.
However, unless she explicitly cancels the implicature (for instance, by
explaining how it is that she doesn't know where Gerard lives despite
being in touch with him), she will be taken to have implicitly admitted
that she misinformed Peter. As this example shows, just as the choice of
276
Postface
words may commit a speaker to unwanted explicatures, so the contextual

implications of an utterance may commit her to unwanted implicatures.
The second type of case where the revised presumption of relevance
leads to better analyses has been much discussed in the literature under
the label 'scalar implicatures'.32 Here is a typical example. In most
situations, the utterance in (19) would implicate (20) or (21):
(19) Some of our neighbours have pets.
(20) Not all of our neighbours have pets.
(21) The speaker doesn't know whether all her,; neighbours have pets.
These implicatures do not always go through, as witness (22) and (23):

(22) Some of our neighbours certainly have pets; maybe they all do.
(23) (a) Peter: Do some of your neighbours have cats, dogs, goldfish,
that sort of thing?

(b) Mary: Yes, some of our neighbour do have pets; in fact they
all do.
At first blush, these facts are reasonably well explained in Gricean terms.
A speaker who knew that all her neighbours have pets and who merely
said, without the sort of qualifications in (22) and (23), that some of her
neighbours have pets, would be giving less information than required by
the first maxim of Quantity. To preserve the assumption that the speaker
is obeying the Gricean maxims, the hearer must take her to implicate
that she doesn't know whether all her neighbours have pets, or more
strongly, that not all them do.
This Gricean account is not without weaknesses. It leaves open the
question of how much information is required on a given occasion by
the first maxim of Quantity - and hence of when 'some' actually carries
an implicature. Nor does it offer any obvious way of deciding when
'some' implicates 'not all' (which it seems to do most of the time), and
when it merely implicates ignorance on the part of the speaker. Still,
'some' conveys 'not all' so often that the implicature from one to the
other is considered by most Griceans (e.g. Levinson 1987) to be a case
of 'generalised implicature', working as a default inference automatically
made, though defeasible in the presence of negative evidence. 33
In the unrevised version of relevance theory, examples like (19) raise
the following problem. Consider a situation where the fact that (at least)
some of the speaker's neighbours have pets would be relevant enough to
be worth the hearer's attention. Then, having recovered this basic
interpretation (on which 'some' is compatible with 'all'), the hearer
would have no reason to go further and assume that the speaker meant
'some, but not all'. This is not an altogether undesirable result. In some
cases it makes the right prediction, as in the following dialogue:
Post/ace
277
(24) (a) Henry: If you or some of your neighbours have pets, you
shouldn't use this pesticide in your garden.
(b) Mary: Thanks. We don't have pets, but some of our neighbours certainly do.
Here, it seems to us, the fact that at lea~t some of Mary's neighbours
have pets is relevant enough, and there is no reason to assume she meant
that not all of them do (or that she doesn't know whether all of them
do). Griceans who treat the inference from 'some' to 'not all' as a
generalised implicature would have to claim that Mary's utterance does
have this implicature, or that the hearer of (24b) would first make this
inference and then (for what reason?) cancel it. Neither hypothesis
seems plausible to us.
However, in some cases the predictions of the unrevised presumption
of relevance are not obviously correct. This happens when the basic
interpretation of 'some' (where 'some' is compatible with 'all') is relevant
enough to be worth the hearer's attention, but when it would clearly be
more relevant to the hearer to know whether 'not all' is the case too. An
example is (25):
(25) (a) Henry: Do all, or at least some, of your neighbours have pets?
(b) Mary: Some of them do.
Here, Henry has made manifest that it would be relevant to him to
know not only whether some of Mary's neighbours have pets, but
whether all of them do. An unrevised relevance model, applied mechanically to this case, would predict that Henry should stop at the first
interpretation that is relevant enough; this is clearly the one on which
Mary is taken to communicate that she has at least some neighbours
who have pets, and nothing more. This prediction is manifestly wrong.
Mary's answer would normally be taken to convey that not all of her
neighbours have pets.
It would, of course, be easy enough to apply the relevance model
flexibly: one might argue, for instance, that someone who asks a
question automatically makes it manifest that what he would consider
relevant enough is nothing less than a full answer to his question, or an
utterance at least as relevant as that. In that case, Mary's answer in (25b),
understood as conveying only that she has at least some neighbours who
have pets, would not be relevant enough. Standard relevance considerations would cause it to be interpreted as implicating34 that not all her
neighbours have pets, thereby satisfying Peter's expectation of adequate
relevance.
However, we much prefer a model that can be applied mechanically.
Isn't this what taking cognitive science seriously is all about? The revised
278
Postface
relevance model is much more satisfactory in this respect (by which we

mean not that we have a full-fledged, mechanically applicable model,
but that at least we don't need to invoke special factors, however
plausible, to account for not-so-special cases). With the revised presumption of relevance, the analysis of example (25) goes as follows.
Mary's answer makes it manifest that she is either unable or unwilling
to inform Peter that all her neighbours have pets. Either implication
would increase the relevance of her utterance. In fact, in most circumstances Mary's answer will make it manifest that she is unable (rather
than unwilling). This inability can in turn be explained in two ways:
either she doesn't know whether all her neighbours have pets, or she
knows that not all of them do. If one of these mutually incompatible
assumptions is manifest enough, it will (in general) be mutually manifest
that Mary intended it to be manifest, since it increases the relevance of
her utterance and is compatible with her preferences. The resulting
interpretation is the one consistent with the principle of relevance.
Mary's answer in (25) is a case where the speaker has deliberately
chosen to express a less informative proposition when a closely related,
equally accessible and more informative proposition would have
demanded no more effort, either from Mary or from the hearer. All
such cases have a similar analysis. If the more informative proposition
would not have been more relevant, there is no implicature. If the more
informative proposition would have been more relevant, the utterance
will be taken to implicate either that the speaker is unwilling, or (more
commonly) that she is unable to provide the more relevant information.
In the latter case, the communicator's inability may be due either to her
not knowing whether the more relevant information is true, or to her
knowing it to be false. If either of these two possibilities is manifest and
relevant, it will be treated as an implicature.
3.4 Far too early to conclude
There are many other aspects of relevance theory that we would like to
see developed, and that we or others have begun working on in articles
and unpublished lectures. Many involve local revisions of the version of
the theory presented in this book. Some open new perspectives that may
turn out to be more important in the general balance of the theory than
the present revisions.
Experimental studies testing relevance-theoretic hypotheses have just
begun, and we hope that they will lead to revisions, new insights, and,
perhaps more important, new problems to investigate. Interesting applications of the theory to literary studies suggest that it might be of some
relevance, more generally, in the study of various cultural productions.
Postface
279
Novel insights and new problems should come from the formal modelling of the theory, possibly with the use of spreading activation models
which seem particularly well suited to representing, on the one hand
the role of accessibility, and, on the other, the way the system'~
computations can be guided on line by monitoring its efforts and effects.
Two important and related domains have hardly been explored at all
from a relevance-theoretic perspective: the theory has been developed
from the point of view of the audience of communicative acts, and
without taking into account the complex sociological factors richly
studied by sociolinguistics. The cognitive processes at work in the
communicator, and the social character and context of communication
are, of course, essential to the wider picture, to the study of which we
hope relevance theory can contribute, and from which it stands greatly
to benefit.
We ourselves have been working on a revised and more detailed
description of inferential comprehension, integrating in particular the
processes involved in enrichment and the comprehension of loose talk
or metaphor. This work will be presented in our forthcoming Relevance
and Meaning.
Notes to First Edition
CHAPTER
COMMUNICATION
1 See Reddy 1979 for a discussion of these misleading metaphors.

2 For instance, by Jakobson (1,960), who uses it as the basis for a classification of
the functions of language, and Lyons (1977), who finds the model incomplete
but not radically inadequate to account for verbal communication.
3 For an examination of the work of Levi-Strauss in this light, see Sperber 1985:
chapter 3.
4 See Smith and Wilson 1979 for a general assessment.
5 For a discussion of semiotic approaches to cultural and artistic symbolism
and suggestions for an alternative approach, see Sperber 1975a, 1980.
6 A rather infelicitous term proposed by C. W. Morris (1938), who defined
syntax as the study of the formal relations among signs, semantics as the study
of the relation between signs and their denotations, and pragmatics as the
study of the relation between signs and their users or interpreters. For a
discussion of the current scope of pragmatics, see Searle, Kiefer and Bierwisch
(eds) 1980: Introduction; Levinson 1983: chapter 1.
7 See Grice 1975, 1978. For a survey of the Gricean pragmatic literature, see
Levinson 1983: chapter 3. Attempts to spell out the Gricean programme in
more detail include Bach and Harnish 1979; Leech 1983. In France, Ducrot
(1972, 1980a, 1980b and a number of other works) has developed a programme
in some ways comparable to Grice's.
8 See Loftus 1979, Neisser 1982.
9 Our discussion in this section applies more directly to Schiffer's version than
to Lewis's. See also note 29 below.
10 For further discussion of the mutual-knowledge hypothesis, see JohnsonLaird 1982a; Sperber and Wilson 1982.
11 See Armstrong 1971; Bach and Harnish 1979; Bennett 1976; Blackburn 1984;
Davidson 1984a; Davies 1981; Grice 1957, 1968, 1969, 1982; Harman 1968;
Lewis 1969; Loar 1976, 1981; McDowell 1980; Patton and Stampe 1969;
Recanati 1979, 1987; Schiffer 1972; Searle 1969, 1983; Strawson 1964a, 1969,
1971; Wright 1975; Yu 1979; Ziff 1967.
282
Notes to First Edition pages 24-31
12 A notable exception being A. H. Gardiner 1932.

13 For ease of exposition, and unless the context indicates otherwise, we will
assume that the communicator is female and the audience male.
14 We are considering here Searle's views as expressed in Speech Acts (1969), not
his more recent and somewhat different views developed in Intentionality
(1983). One of Searle's arguments was in the form of a purported counterexample to Grice's analysis (Searle 1965: 221-39; 1969: 44-5). This counterexample has been satisfactorily dealt with by Grice 1968: 160-5; and Schiffer
1972: 27-30.
15 For other examples of codeless communication, discussed in greater detail, see
.
section 10.
16 Note that it would not be too difficult to reconcile the strong inferential theory
of communication with a modified code theory. The code theorist might
concede to the inferential theorist that all codes are sets of conventions and that
decoding is an inferential process along the lines described above, and the
inferential theorist might concede in return that the inferences involved in
communication are decoding inferences. However, the resulting compromise
would combine the worst defects of both theories: it would fail to take into
account the role of uncoded inference in communication, and it would ignore
the non-inferential character of much decoding.
17 We assume that the 'response' involved in intention (27a) is always that the
audience should be informed of something (in the broad sense in which we are
using 'inform'). This is by no means a commonly accepted view. Grice himself
initially had in mind two types of response: belief, in response to a statement,
and action, in response to an injunction. He later (Grice 1968, 1969) excluded
action in response to an injunction and considered only mental responses: in
the case of a statement, the intended response is the recognition by the
audience that the communicator has a certain belief and, sometimes, the
adoption of the same belief; in the case of an injunction, the intended response
is the recognition by the audience that the communicator has a certain
intention, and the adoption of the same intention. Others have expressed still
different views (see Searle 1969; Armstrong 1971; Bennett 1976). We develop
our own view in sections 8 to 12, and more specifically in section 11.
18 For examples and discussion suggesting that this intention is unnecessary, see
Schiffer 1972: chapter 3.
19 See for instance Grice 1982; Recanati 1987.
20 One may also, as suggested by Searle (1969: 47), replace the infinity of
intentions by a reflexive intention, i.e. a complex intention comprising as one
of its sub-intentions the intention that the whole complex intention should be
recognised. Such a reflexive intention is 'overt' in much the same way as an
ordered infinity of intentions. It might seem that a single reflexive intention is
psychologically more plausible than an infinity of intentions, but we doubt
this for the following reason. Normally, when a representation contains a
definite reference to a representation, this reference can be replaced by a
mention of the representation referred to. For instance (a) contains a reference
to the representation expressed by Mary and spelled out in (b); hence (c) can
validly be inferred from (a)-(b):
283
(a) Peter believes what Mary said.

(b) Mary said that it is raining.
(c) Hence: Peter believes that it is raining.
21
22
23
24
25
26
27
28
Often, understanding a representation such as (a), or establishing its relevance,

involves just such a substitution. A case in point is a communicator's intention
I that her audience should recognise her intention]: intention I cannot be
fulfilled or fully grasped without grasping]. With a reflexive intention J which
includes the sub-intention that the audience recognise J, this yields an
infinitely long formula. Since infinitely long formulas are not available, let
alone intelligible, to the mind, it yields a headache. Reflexive intentions do not,
after all, provide a psychologically plausible way of making the notion of
'overtness' more precise.
Though the infinity-of-intentions proposal (and its reflexive variant) on the
one hand, and the mutual-knowledge proposal on the other, both dispose of
examples of the type suggested by Strawson 1964a and developed by Schiffer
1972, it is possible to think of other examples of untypical communication
which are handled differently by the two approaches. The infinity-ofintentions approach does not rule out cases where the communicator's
intentions, though recognised by the audience, do not become mutually
known. The mutual-knowledge approach does not rule out cases where
mutual knowledge, and hence communication, is established unintentionally,
or at least seems so to the audience. These cases have not been discussed in the
literature. Since we are not aware that any definite conclusion against one or
the other solution follows from these extra cases, we leave it at that.
A notable exception being the psychologist Herbert Clark and his associates.
See Clark 1977, 1978; Clark and Lucy 1975; Clark and Haviland 1977; Clark
and Marshall1981; Clark and Schunk 1980; Clark and Carlson 1981.
The William lames Lectures, Logic and conversation, delivered at Harvard in
1967, brought together ideas first put forward in Grice 1957 and 1961.
Versions of the second and third lectures were published as Grice 1975, 1978;
parts of the fourth lecture are summarised in Grice 1981; many of the ideas in
the last three lectures were presented" in Grice 1968, 1969; the complete text,
with an important 'Retrospective Epilogue', was published in Grice 1989.
The role of the Gricean maxims in disambiguation was not discussed by Grice
himself, but was dealt with by Katz 1972: 449-50; Walker 1975: 156-7, Wilson
and Sperber 1981: 156-9.
A temptation to which Leech 1983 might be felt to have yielded.
We have tried (Wilson and Sperber 1981) to show that all the maxims can be
reduced to a single well-defined maxim of relevance.
Kempson 1975, Stalnaker 1974, Wilson 1975 de"elop Griceap. accounts of
presuppositional phenomena; Sadock 1979, Levinson 1983 (147-62) look at
metaphor and figurative language from a Gricean perspective; Searle 1975,
Bach and Harnish 1979 approach indirect speech acts in Gricean terms. See
also note 7 above.
It has been argued (see Fodor 1983: 102) that conceptual identifications of
distal stimuli, which are the output of perceptual processes, need inferential
284
29
30
31
32
validation before being accepted by the individual as facts. If this were so,
being perceptible would not be a sufficient condition for being manifest. The
best evidence for this claim is the fact that one can mistrust one's senses, and
hence perceive and yet not believe. However, one can also infer and not
believe, as when a validly inferred conclusion contradicts a strongly held
belief. It seems to us that the output of perception, just like that of inference,
requires no validation in order to be accepted as true. On the other hand, the
output of perception (like the output of inference) can be inferentially
invalidated. To be more precise, therefore, we might say that to be manifest is
to be capable of being perceived or inferred without being immediately
invalidated.'
What, for instance, Lewis (1969: 56) calls a basis for common (i.e. mutual)
knowledge is roughly equivalent to our mutual manifestness. We part
company with him when he goes on to state, as a mere matter of definition,
that the existence of such a basis is a sufficient condition for the existence of
common knowledge itself. See also Clark and Marshall1981.
It would be unfair not to mention that inferential theorists generally resist the
temptation. Grice (1982: 237), who develops an inferential account of how
language could have originated, calls it 'a myth'. See also Lewis 1975/1983:
181.
See Catherine Kerbrat-Orecchioni 1977 for a survey.
This is why we did not analyse the informative intention as an intention to
make I mutually manifest. This would not be adequate in the cases where
the communicator does not herself believe the information she is trying to
communicate. There is another way of handling this problem, though:
ostensive communication could be described as an attempt to create a
genuinely mutual cognitive environment between social personae. When the
communicator is sincere (and so is the audience in manifesting its acceptance of
the information communicated), then the actual individuals and their social
personae coincide, and otherwise they don't. This formulation, which is a
notational variant rather than a substantive alternative to the one we follow
here, might be more appealing from a sociological point of view.
CHAPTER
INFERENCE
1 Not all conclusions implied by a set of premises can be generated by inference

rules alone. For instance, a premise P implies an infinity of conclusions of the
form (P or Q), where Q is any assumption whatsoever, and some noninferential means of generating Q would be required. But those conclusions
which cannot be generated are cognitively uninteresting; they are 'trivial' in a
sense to be discussed below in section 5.
2 Incidentally, there is a paradox in insisting that verbal communication is
constrained by the mutual-knowledge requirement while at the same time
recognising that non-demonstrative inference is involved. The point of the
mutual-knowledge requirement is to make it possible to account for verbal
communication by means of a failsafe algorithm; to recognise the role of
non-demonstrative inference is to rule out the possibility of such an algorithm.
285
If the paradox is not more blatant, it is, we fear, because of the general haziness
with which these issues are usually discussed.
3 Here we assume only token identity between brain states and mental states.
See Fodor 1974.
4 For the role of incomplete logical forms in speculative thinking, see Sperber
1985: chapter 2.
5 Or storage format: the important point being that all representations stored in
that store or format are retrievable and processable in the same way, and
differently from representations stored otherwise.
6 See Sperber (1985: chapter 2) for elaboration and discussion of this distinction.
7 Because logical forms, propositional forms and factual assumptions are not
directly observable, we will have to use natural-language sentences to
represent them, despite the lack of any one-to-one correspondence between
sentences on the one hand and logical forms, propositional forms and factual
assumptions on the other. In practice, this should present no more problem
than it does in everyday communication, when hearers or readers normally
have no difficulty in identifying the assumptioI1-a given utterance was intended
to express. We do not mean to imply that natural language reflects the
structure of the language of thought more closely than is required by the fact
that both are formal objects with semantic properties, and that the one can be
successfully used to communicate the other.
8 And, for that matter, even when he is capable of doing so consciously, as
evidenced by the work of Kahneman and Tversky.
9 See Kahneman, Slovic and Tversky 1982: especially chapters 1, 34 and 35.
10 In what format are assumption schemas stored in the mind? They could be
embedded in factual assumptions which state that some completion of the
schema is or might be true. Or they could be stored as fully propositional
factual assumptions, but with very weak empirical import, which achieve
relevance only when strengthened by the addition of new constituents. For
instance, assumption schema (29) below could be stored as (i) or as (ii):
. (29) The outside temperature is - - degrees centigrade.
(i) For some number n, 'the outside temperature is n degrees centigrade'
is true.
(ii) The outside temperature is some number of degrees centigrade.
11
12
13
14
Since we have no principled argument in favour of either format, or of any of

the other formats which could be imagined, we will not pursue this issue.
This is not to say that no such system will ever be developed: JohnsQn-Laird
1983 outlines a research programme designed to develop an alternative to
models of inference based on deductive rules.
This fits well with the general view of cognitive systems as purely computational that is currently being developed in cognitive psychology. See Fodor
1980.
See, for example, Katz 1972; Fodor 1981 a; Fodor, Garrett, Walker and Parkes
1980.
For discussion of these notions, see Winograd 1977; Minsky 1977; and Schank
and Abelson 1977.
286
15 Kripke 1972; Putnam 1975; see also Fodor 1982.

16 For survey and discussion, see Pulman 1983; Carston 1984a.
17 We have only considered the propositional content of encyclopaedic entries;
there is no reason, however, why they should not contain - or give access to'images' and whatever types of mental object can be used as sources of
information in conceptual thinking.
18 Fodor, Garrett, Walker and Parkes 1980, Fodor 1981a, reacting against classic
decompositionalist accounts such as Katz 1972, Miller lnd Johnson-Laird
1976. For survey and evaluation of the decompositionalist approach, see J. D.
Fodor 1977.
19 For interesting discussion of these questions, see Carston (in preparation).
20 For convenience, we will describe the working of the deductive device for
assumptions; it works in exactly the same way for all other logical forms.
21 See Rips 1983. This also contains an excellent survey of previous work on the
psychology of natural deduction.
22 For discussion and justification of the distinction between 'forwards' and
'backwards' rules, see Rips 1983.
23 And, possibly, to optional constraints inhibiting parts of the derivation which
are unlikely to contribute to the pursuit of relevance; see chapter 4, section 5.
24 We are using this hypothetical rule for expository purposes, and without
intending thereby to suggest that such a rule exists. It is quite conceivable that
the relation of containment, and other transitive relations as well, are handled
not by deductive rules, but by mental models a la Johnson-Laird.
25 In general, new information is not necessarily information that is new to the
organism, but merely information that is being newly processed. Information
retrieved from memory could be new information in this broader sense. The
organism must have some rationale for deciding, for any newly retrieved item
of information P, whether it is best added to the context in which some other
information Q is being processed, or whether to treat P as new information
and process it in a context containing Q. Considerations of relevance should
weigh heavily here, as in other aspects of cognition.
26 We might also give a more formal characterisation of the conditions under
which a contextualisation has contextual effects: Let C be a context and P a
set of new premises. Let Conclusions ofP be the set of conclusions deducible
from P alone, Conclusions of C the set of conclusions deducible from C
alone, and Conclusions of P U C the set of conclusions deducible from the
union of P and C. Let two assumptions with the same content but with
different strengths count as two different assumptions. Then the contextualisation of P in C has no contextual effect if and only if the two following
conditions are met:
(i) Conclusions ofC is a subset of Conclusions ofP VC;
(ii) the complement of Conclusions ofC with respect to Conclusions of pvc
is a subset of Conclusions ofP.
H conditions (i) and (ii) are not both met, then the contextualisation of P in
C has some contextual effect.

CHAPTER
287
RELEVANCE
1 For the notion of a comparative concept, see above, chapter 2, section 3.

2 As suggested in chapter 2, note 24, it is quite conceivable that at least some of
the inferences determined by transitive relationships such as richer than are
computed not by use of inference rules but by use of 'mental models'. Even if
this is so, effort is involved, effects are achieved, and the notion of relevance
applies as we have described.
3 A view once attributed to us by Gazdar and Good 1982.
4 See for example Brown and Yule 1983: chapter 2; Levinson 1983: chapter 1.4;
Lyons 1977: chapter 14.
5 This is a complication which could be accommodated in a ~traightforward way
within the framework we are proposing, but to do so would involve too much
effort for too little effect.
6 Remarks along these lines are made by Johnson-Laird 1967, Stenning 1978,
Stalnaker 1978, McCawley 1979, Sag 1981 and most notably by Hobbs 1979.
However, what many of these authors have in mind is a restricted subset of
implicated contextual assumptions (often analysed as 'pragmatic presuppositions'), rather than the full set of contextual assumptions used in the
interpretation of an utterance.
7 We may assume that the memory of the deductive device has a limited, indeed
a rather small capacity, so that no extensions beyond that capacity are possible.
The maximal contexts accessible are therefore those which, in view of their
size, cannot be extended further.
8 In some contexts, say in a study of different processing strategies for the same
stimulus, it might be desirable to compare the relevance of the same
phenomenon relative to different possible methods of processing. The
comparative definition of the relevance of a phenomenon could be adapted for
this purpose in an obvious way.
9 And not, of course, in the sense of speech-act theory.
10 It is true that Grice's theory is put forward as a theory of 'conversation'.
However, it has invariably been taken as a more general theory of verbal
communication, and Grice has done nothing to correct that interpretation.
11 Earlier versions of relevance theory were closer in these respects to Grice's
approach. We had generally assumed that there was a presumption of maximal
rather than optimal relevance, and that communicator and audience had to
have and use knowledge of the principle of relevance. However, the idea that
the principle was exceptionless was there from the start.
CHAPTER
ASPECTS OF VERBAL COMMUNICATION
1 See Marslen-Wilson 1973; Fodor 1983: 61-64.

2 For example Bach and Harnish (1979: 7) assume that there is a 'linguistic
presumption' and a 'communicative presumption' which together have the
desired effect.
288
3 See for example Bach and Harnish 1979: 20-23.

4 See Carston 1984b, 1988a for a reanalysis of the temporal 'implicatures' of
'and' as aspects of explicit content within a relevance-based framework.
Kempson 1986 and Carston 1988a propose reanalyses of 'scalar' or 'quantity
implicatures'. Cormack 1980 and T ravis 1981, 1985 have independently
suggested that not all aspects of explicit content are strictly linguistically
determined. See also Blakemore 1987, 1988a.
5 The example comes from Winograd 1977.
6 See for example Swinney 1979, Hogaboam and Perfetti 1975, Tyler and
Marslen-Wilson 1977, Cairns and Kamerman 1975, Tanenhaus and Lewman
1979, and more generally Marslen-Wilson and Tyl~r 1980 and Fodor 1983.
7 It is arguable that the referents of at least some proper names, e.g. 'Mt Everest',
'The Eiffel Tower', are recoverable by decoding alone. The greater the range of
possible referents, the less appealing pure decoding solutions become.
8 We do not want to deny that there might be what Quine calls 'eternal
sentences'. An eternal sentence, as he defines it, is 'a sentence whose truth
value stays fixed through time and from speaker to speaker' (Qui ne 1960: 193).
He gives the following (actually less than perfect) example: 'Copper oxide is
green' (ibid: 12). An eternal sentence, if used literally, expresses the same
proposition in any context, and therefore leads two people who understand
that sentence to entertain exactly the same thought. What we do very much
doubt is that for every thought there is a corresponding eternal sentence. The
fact that some sentences correspond to a single thought is hardly more
significant than the fact that some random strings of fifty letters and spaces
correspond to a sentence of English. Quite generally, a single sentence, or even
a single sense of a sentence, does not correspond to a single thought, and a
single thought does not correspond to a single sentence.
9 See for example Green and Morgan 1981: 170-71; Clark 1977: 420.
10 A similar point is made by Harnish (1976: 346), which raises many interesting
questions about Grice's account of implicatures; see also Walker 1975, Hugly
and Sayward 1979, Sadock 1978 and Wilson and Sperber 1986a.
11 We are using capitalisation to represent both focal (sentential, nuclear) and
contrastive stress. For reasons of space, we will largely ignore the effect of
secondary stress and tone group on utterance interpretation.
12 For survey and discussion from rather different perspectives, see Rochemont
forthcoming; T aglicht 1984: chapters 1-3; Brown and Yule 1983: chapters
3-5; Reinhart 1981; Prince 1981; Givon (ed.) 1979; Oh and Dinneen (eds)
1979; Clark and Haviland 1977; Lyons 1977: chapter 12.7; Chafe 1976;
Jackendoff 1972; Halliday 1967/8.
13 For survey and discussion see Johnson-Laird 1983: chapter 13.
14 For survey and discussion see note 6 above.
15 See for example Rochemont forthcoming; Allerton and Cruttenden 1979.
16 Nor need it prevent these natural functions being taken over by some purely
linguistic device, syntactic, morphological or intonational: indeed the costlier
it is to vary the position of stress in a language, the more one would expect to
find such things as focusing particles to compensate.
17 See for example Gussenhoven 1983; Rochemont forthcoming. If an approach
18
19
20
21
22
23
24
25
289
along the lines suggested here could be worked out, a more general
consequence would be that there would be no need for a notion of focus in
generative grammar. The issues are complicated and have been the subject of
some sophisticated syntactic argumentation. What is worrying about this
argumentation is that for all its complex machinery it often ends in appeals to
pragmatic notions whose nature is left unspecified. We have simply tried to
sketch in the lines of a pragmatic account which should serve at least as a
supplement to the linguistic notion of focus, and at best as a replacement for it.
For discussion, see Wilson 1975; Kempson 1975; Gazdar 1979; J. D. Fodor
1979; and Soames 1979.
In the same way, it can be shown that cohesion and coherence are derivative
categories, ultimately derivable from relevance. For detailed arguments for
this position, see Blass 1986, 1990.
Note also a technical difference of some consequence: while new or focused
information has invariably been treated as non-propositional, or of less than
propositional size, foreground information is, by our definition, propositional: a foreground implication is an analytic implication, not an NP or VP, or an
NP intension or a VP intension.
This is not to say, however, that no arbitrary linkages between linguistic form
and pragmatic interpretation exist. It used to be suggested in the presuppositionalliterature (e.g. in Stalnaker 1974: 212) that there were certain linguistic
structures whose function was to impose constraints on the contexts in which
utterances containing those structures could occur. Pending an account of the
role of context in utterance interpretation, it was hard to see why such
structures should exist. Some years ago, however, Diane Blakemore suggested
that in a relevance-based framework they might have a significant advantage
from a processing point of view. As we have seen, the speaker can use the
linguistic form of an utterance to guide the interpretation process. Blakemore's idea was that, just as the natural links between intonational structure
and pragmatic interpretation may become grammaticalised, so a language
might develop certain structures whose sole function was to guide the
interpretation process by stipulating certain properties of context and
contextual effects. Clearly, in a relevance-based framework the use of such
structures might be highly cost-efficient. This approach appears to shed light
on a wide range of apparently disparate phenomena on the borderlines
of grammar and pragmatics, and seems to us a particularly promising area
for future research. For a detailed development of this approach, see
Brockway 1981,1983; Blakemore 1985, 1987, 1988a. For further interesting
work along these lines see MacLaran 1982; Kempson 1984; Smith 1983; and
Blass 1990.
The distinction between descriptions and interpretations was developed in
another context in Sperber 1985: chapter 2.
See Searle 1969: chapter 3; Lewis 1975; Bach and Harnish 1979: 10-12,
127-13 I.
Or in some other given world, in the case of fiction for instance.
In an earlier paper (Sperber and Wilson 1981), we analysed irony and free
indirect speech as varieties of mention. We distinguished direct speech,
290
26
27
28
29
30
involving mention of linguistic form, from indirect speech, which we

proposed to analyse as involving mention of logical form; we then assimilated
irony to the case of indirect speech. The problem with this proposal is that the
notion of mention does not really stretch to cover the full range of cases we are
now proposing to handle. Mention is a self-referential or self-representational
use of language: it requires full linguistic or logical identity between
representation and original. One of the implications of section 7 has been that
mention is only a special case of a much more general phenomenon: the use of
a propositional form to represent not itself but some other propositional form
it more or less closely resembles. We have therefore abandoned the term
'mention' in favour of the more general term 'intefpretation'.
Apart from this terminological revision, our account of irony has not
substantially changed. It has been criticised on a number of counts (Kerb ratOrecchioni 1981; Clark and Gerrig 1984); some of the criticisms are answered
in Sperber 1984. The theory has received some experimental confirmation in
Jorgensen, Miller and Sperber 1984.
This is not to deny that some people are better than others at producing or
understanding metaphor or irony; but then some people are better than others
at producing and understanding strictly literal utterances. What is involved
here is not competence but talent.
For excellent accounts of the speech-act approach to pragmatics, see Bach and
Harnish 1979; Recanati 1987. For a more semantic approach to these
phenomena, see Katz 1977.
We take it that promising is different from merely asserting that one will do
something that the hearer wants one to do. In the latter case, someone who
fails to do what she said she would and is accused of not keeping her promise
would be quick to deny having promised, and would be quite right to do so. A
promise is a particular, culturally defined form of commitment. Similarly,
thanking is a particular, culturally defined form of expression of gratitude.
Many societies have other forms of commitment, more akin to swearing, for
instance, and other forms of expressing gratitude, more akin to blessing, for
instance, than the typically modern Western promising and thanking. We have
no doubt that a cross-cultural study of such speech acts would confirm their
cultural specificity and institutional nature.
For an interesting cross-linguistic survey of syntactic sentence types, see
Sadock and Zwicky 1985.
The strength of the assumption explicated also follows from the principle of
relevance. Compare the ordinary assertions (i) and (ii):
(i) My name is Janet.
(ii) Belle Etoile will win the 3 p.m. race.
In normal circumstances, the explicature of (i) would be much stronger than
that of (ii), which would typically have the force of nothing more than an
educated guess. The first part of the presumption of relevance amounts to a
guarantee that the information communicated is relevant enough to be worth
the hearer's attention. Whereas in (i) the assumption that the speaker is
Notes to First Edition page 249
291
anything less than certain of her own name would normally be inconsistent
with manifest assumptions, with (ii), by contrast, an educated guess would be
quite relevant enough
Notes to Second Edition
CHAPTER I
COMMUNICATION
a Our rejection of mutual knowledge in favour of mutual manifestness has

been much discussed (see the commentaries by Bach and Harnish, Gibbs,
Russell, McCawley, Gerrig and Hinkelman in The Behavioral and Brain
Sciences 10.4, and our reply in Sperber and Wilson 1987b). In response to
continuing commentary by Garnham and Perner (The Behavioral and Brain
Sciences 1990: 178-9), we tried in Sperber and Wilson 1990a to bring together
the rather scattered arguments in Relevance and present the differences
between mutual knowledge and mutual manifestness in more perspicuous
terms.
b Our use of the term 'guarantee of relevance' has sometimes been misunderstood as implying that utterances cannot but be relevant. As this passage
makes clear, we are no more committed to the view that utterances cannot
but be relevant than to the view that assertions cannot but be true.
c Throughout this book, the term 'principle of relevance' refers to the (communicative) principle that every act of ostensive communication creates a
presumption of relevance, rather than the more general (cognitive) principle
that human cognition tends to be geared to the maximisation of relevance. In
the Postface, we will suggest that two principles of relevance should be
distinguished: a First (or Cognitive) Principle and a Second (or Communicative) Principle.
CHAPTER 2
INFERENCE
a We would not now assume such asharp distinction between input (specialised) and central (unspecialised) systems. In the last ten years, there has been
growing evidence that so-called central systems should be analysed in
modular terms. See Sperber 1994b for discussion.
b For reservations about the treatment of 'central' systems in the Fodorian
framework, see note a above.
c It is worth emphasising that the deductive system presented in this chapter is
designed to do no more than illustrate one way in which deductive inferences
Notes to Second Edition pages 85-241
294
might be performed. We are a long way from having the sort of evidence that
would choose between the huge range of conceivable methods of performing
deductive inference.
d It is sometimes suggested that we have overlooked a fourth type of contextual
effect, namely weakening of existing assumptions. Weakenings are allowed
for in our formal definition of the conditions under which a contextualisation
has contextual effects (chapter 2, note 26). We assume, though, that weakening is always a by-product of a more basic contextual effect: for example,
contradiction and elimination of an existing assumption weakens all contextual implications which depended on that assumption for some support.
CHAPTER
RELEVANCE
a Definitions (42) and (43) of relevance to an individual (and definitions (58)

and (59) of relevance of a phenomenon in section 6 below) are discussed and
modified in the Postface.
b See chapter 1, note b.
c The presumption of optimal relevance is discussed and modified in the
Postface.
d For further discussion of the relation between co-operation and communication, see the Postface. See also Sperber 1994a.
e By 'consistent with the principle of relevance', we mean consistent with the
particular instantiation of it communicated on that occasion. See p. 162 above
and our reply to Morgan and Green in Sperber and Wilson 1987b: 745.
CHAPTER
ASPECTS OF VERBAL COMMUNICATION
a This definition is too strong as it stands. It should be modified to accommodate the fact that someone who says, for example, 'I tell you that P', or 'P
despite Q', can explicitly communicate P.
b That is, given the particular instantiation of the principle of relevance
communicated on that occasion.
c In preparing the French translation of Relevance, we discovered that, in fact,
this classic example of irony (discussed by Booth 1974: 10) is not the work of
Voltaire but of his English translator. A closer (and duller) translation of the
French original would go: 'Finally, while the two kings had Te Deums sung
in their respective camps .. .'.
Notes to Postface
4
5
6
7
French (Editions de Minuit, Paris, 1989); Russian (part-translation, Progress

Publishers, Moscow, 1989); Japanese (Kenkyusha Shuppan, Tokyo, 1993);
Korean (Hanshin Publishing Co, Seoul, 1993); Italian (Edizioni Anabasi,
Milan, 1993); Spanish (Visor, Madrid, 1994); Bahasa Malaysia (to appear).
This review is far from complete. We have excluded references to working
papers, and listed only selected titles for several authors whose contributions
we would have liked to acknowledge more fully. Important areas of research,
e.g. on style and intonation, have not been given their due. We have done
no more than sample the existing literature, and regret the many omissions
due to ignorance, oversight or lack of space.
A multiple review of Relevance appeared in The Behavioral and Brain
Sciences (10.4, 1987), with continuing commentary by Politzer 1990, Garnham and Perner 1990 and Chiappe and Kukla forthcoming. Major reviews
include Fowler 1989; Hirst 1989; Jayez 1986; Leslie 1989; Levinson 1989;
Mey and Talbot 1988; Pateman 1986; Seuren 1987; Travis 1990; Walker
1989. See also note 8.
See, for example, Sperber and Wilson 1987b, 1990a, forthcoming a; Wilson
1992a; Wilson and Sperber 1987. See also Blakemore 1994a.
Sperber and Wilson 1987a.
Blakemore 1992; Sinclair and Winckler 1991; Moeschler and Reboul1994.
See, for example, Gutt 1986; Smith and Wilson 1992; Sperber 1994a; Wilson
1994a; Wilson and Sperber 1986b, c. Encyclopaedia articles discussing
aspects of relevance theory include Blakemore 1988b, forthcoming; Carston
1988b, 1993a, b; Kempson 1988b; Leech and Thomas 1990; Moeschler and
Reboul 1994.
For critique and discussion, see note 3 above, and Berg 1991; BurtonRoberts 1985; Chametzky 1992; Charolles 1990; Culpeper 1994; Escandell
Vidal 1993; Gibbs 1987; Giora 1988; Gorayska and Lindsey 1993; Grundy
1995; Nebeska 1991; Nishiyama 1992, 1993, 1995; O'Neill 1988; Roberts
1991; Sadock 1986; Sanchez de Zavala 1990; Sinclair 1995; Sun 1993; Taylor
and Cameron 1987; Toolan 1992; Wilks and Cunningham 1986; Ziv 1988.
For replies, see note 4 above.
296
Notes to Postface pages 256-257
9 Published dissertations include Blakemore 1987; Blass 1990; Forceville

1994a; Gutt 1991; Perrin (forthcoming); Tanaka 1994; Vandepitte 1993.
Unpublished dissertations include Austin 1989; Campbell1990; Clark 1991;
Espinal1985; Ferrar 1993; Groefsema 1992a; Happe 1992; Ifantidou 1994;
Itani 1995; Jodlowiec 1991; Mao 1992; Matsui 1995; Pilkington 1994;
Politzer 1993; Posnanski 1992; Reboul 1990a; Rouchota 1994a; Stainton
1993; Zegarac 1991. Monographs include Moeschler 1989b; Nasta 1991;
Reboul 1992a.
10 These include Carston et al. forthcoming; Davis 1991; Guijarro Morales
1993; Kasher forthcoming; Kempson 1988a; Moeschler 1989c, Moeschler et
al. 1994; Smith 1989; two special issues of Linguae (Wilson and Smith 1992,
1993) which contain a representative collection of papers; and the annual
University College London Working Papers in Linguistics, which contain a
wide variety of interesting work not all of which can be individually
mentioned here.
11 Mitsunobu 1993.
12 An interesting reassessment of Grice's contribution to the philosophy of
language can be found in Neale 1992. Neale cites a passage from Grice
(1957: 222) where it is suggested that considerations of relevance can help
with disambiguation.
13 This included studies of reference assignment within a broadly Gricean
framework by the psycholinguist Herb Clark and his associates (see e.g.
Clark 1977, Clark and Haviland 1977, Clark and Marshall1981), in which
the notion of a 'bridging implicature' was introduced. The treatment of
bridging implicatures within the relevance-theoretic framework is discussed
by Matsui 1993, 1995; Wilson 1992b, 1994b; Wilson and Sperber 1986a. See
also note 15.
14 Bach and Harnish 1979, an excellent Gricean account of communication,
abandons the Co-operative Principle and maxims when it comes to disambiguation, switching instead to informal talk of 'contextual appropriateness'.
Levinson 1983, the standard textbook on pragmatics, contains nothing on
disambiguation, but discusses the role of 'implicature' in reference assignment on pp. 34-5. More recently, Levinson (1987, 1988) has done important
work on reference assignment in a neo-Gricean framework, where he
continues to talk of 'implicatures' as contributing to truth conditions not
merely in bridging but in reference assignment in general, and in disambiguation too.
15 Various aspects of the relevance-theoretic treatment of reference assignment
are discussed in Ariel1990; Blass 1986; Forget 1989; Foster-Cohen 1994;
Fretheim forthcoming a; Gundel forthcoming; Hawkins 1991; Kempson
1988c; Kleiber 1990, 1992; Matsui 1993, 1995; Reboul 1992b, 1994a,
forthcoming; Recanati 1993; Rouchota 1992, 1994a, d; Wilson 1992b, 1994b.
16 See, for example, Carston 1988a, 1993c, forthcoming b. For discussion, see
Atlas 1989; Bach 1994a, b; Levinson 1987, 1988; Neale 1992; Recanati 1989;
Wilson and Sperber 1993, forthcoming.
17 On enrichment, see, for example, Atlas 1989; Bach 1994a, b; Bertolet 1990;
Bertuccelli-Papi 1992; Blakemore 1989a; Espinal 1993; Groefsema 1995,
18
19
20
21
22
23
24
25
26
27
28
297
forthcoming; Haegeman 1987, 1989; Hirst 1987; Horn 1992; Kandolf 1993'
Klinge 1:93; Moeschler 1993b; Recanati 1994, forthcoming; Scancarelli
1986; Stamton 1993, 1994; Taylor 1993; and references in note 16.
See, for example, Blakemore 1987; Carston 1988a.
See Grice 1989: 121-2,361-3.
For survey and discussion of early work on non-truth-conditional semantics
see Wilson 1975. For important work outside the relevance-theor~tic frame~
work, see Ducrot 1980b, 1983, 1984, Ducrot et al. 1980.
For procedural accounts of discourse connectives, see Ariel1988; Blakemore
1988a, b, 1990, 1993; Blass 1990, 1993; Ducrot 1984, Ducrot et al. 1980;
Gutt 1988; Haegeman 1993; Higashimori 1992a, b, 1994; Itani 1995; Jucker
1993; Luscher 1994; Moeschler 1989a, b, 1993a; Smith and Smith 1988;
Unger 1994; Vandepitte 1993; Wilson and Sperber 1993. For related accounts
of procedural semantics, see Gabbay and Kempson 1991; Jiang 1994;
Kempson forthcoming.
Relevance-theoretic accounts of mood indicators are developed in Clark
1991, 1993a, b; Lunn 1989; Rouchota 1994a, b, c; Wilson and Sperber
1988a, b, 1993. For discussion of various aspects of the relevance-theoretic
approach to speech acts, see Bird 1994; Clark 1991; Groefsema 1992b;
Harnish 1994; Moeschler 1991; Reboul 1990b, 1994b; Recanati 1987; see
also Kasher 1994.
On mood indicators, see note 22. On discourse particles and adverbials, see
Blass 1989, 1990; Espinal1991; Ifantidou 1994; Ifantidou-Trouki 1993; Itani
1995; Konig 1991a, b; N0lke 1990; Watts 1988; Wilson and Sperber 1993;
Yoshimura 1993b. On parentheticals, see Blakemore 1990/1; Espinal 1991;
Ifantidou 1994; Wilson and Sperber 1993. On tense and aspect, see Moeschler 1993b; Smith 1993; Zegarac 1991, 1993.
For development and applications of the relevance-theoretic accounts of
metaphor and irony, see Forceville 1994a, b; Hymes 1987; Pilkington 1992,
1994; Reboul 1990a, 1992a; Song forthcoming; Sperber and Wilson 1985/6,
1990b; Wilson and Sperber 1988b, 1992; Vicente 1992; Yoshimura 1993a.
For discussion, see Gibbs 1994; Goatly 1994; Hamamoto forthcoming;
Kreuz and Glucksberg 1989; Martin 1992; Perrin forthcoming; Recanati
forthcoming; Seto forthcoming.
See Gutt 1990, 1991, 1992; Tirkkonen-Condit 1992; Winckler and van der
Merwe 1993.
On echo questions, see Blakemore 1994b; on reformulations, see Blakemore
1993; on pseudo-imperatives, see Clark 1991, 1993a; on hearsay particles,
see Blass 1989, 1990, Ifantidou 1994, Itani 1995; on metalinguistic negation,
see Carston forthcoming a, Moeschler 1992; see also Burton-Roberts
1989a, b; Fretheim forthcoming b; Yoshimura 1993b. For interesting application of the related notion of 'polyphonie', see for example Ducrot 1983.
The implications of relevance theory for literature are discussed in Durant
and Fabb 1990; Fabb forthcoming, in preparation; Green 1993; Kiparsky
1987; Pilkington 1991, 1992, 1994; Reboul 1990a, 1992a; Richards 1985;
Sperber and Wilson 1987b: 751; Trotter 1992; Uchida forthcoming.
On humour, see Ferrar 1993, Jodlowiec 1991; on politeness, see Austin
298
29
30
31
32
33
34
1989, Jucker 1988; on advertising, see Forceville 1994a, b, Tanaka 1992,

1994; on argumentation, see Campbell 1990, 1992, Moeschler 1989b, c; on
political language, see Wilson, J. 1990; on language in education, see Mayher
1990; on cinema, see N asta 1991.
On the implications of relevance theory for linguistics, and steps towards
formalisation, see Gabbay and Kempson 1991; Kempson 1988c, forthcoming; Jiang 1994; see also Posnanski 1992.
True, Grice's maxim of Relation is just 'be relevant.' His two maxims of
Quantity, however, suggest the maximisation of relevance in our sense. The
first maxim of Quantity (,Make your contribution as informative as is
required') goes towards increasing effect. The secdnd maxim of Quantity
(,Do not make your contribution more informative than is required') goes
towards minimising effort (as do the maxims of Manner). See Horn 1984,
1988; Levinson 1987, 1988 for discussion.
This was noted in Relevance in the text surrounding the presumption of
relevance, but was not built into the presumption itself.
See Carston 1988a, forthcoming b; Harnish 1976; Horn 1984, 1988; Levinson 1987, 1988.
See Carston (forthcoming b) for discussion of Levinson's approach from a
relevance-theoretic perspective.
For expository purposes, we do not question here the standard view that the
richer overall interpretation is reached via an implicature. However, as noted
above, there is evidence that at least some cases of so-called 'quantity
implicature' are cases of enrichment instead.
Bibliography
Allerton, D. and Cruttenden, A. (1979), 'Three reasons for accenting a definite

subject'. Journal of Linguistics 15.1: 49-53.
Alves, H. O. (ed.) (1986), Encontro de Linguistas: Acta. Universidade do
Minho, Minho, Portugal.
Anzai, Y. et al. (eds) (1992), Ninchi-kagaku Hando-bukku (Handbook of
cognitive science). Kyoritsu Publishing, Tokyo.
Ariel, M. (1988), 'Retrieving propositions from context: Why and how'. Journal
of Pragmatics 12.5/6: 567-600.
Ariel, M. (1990), Accessing noun-phrase antecedents. Routledge, London.
Aristotle (1963), De interpretatione, translated by]. L. Ackrill, Clarendon
Aristotle Series. Oxford University Press, Oxford.
Armstrong, D. (1971), 'Meaning and communication'. Philosophical Review,
80: 427-47.
Arnauld, A. and Lancelot, C. (1968), Grammaire de Port-Royal, English
translation edited by R. Alston. Scolar Press, Menston, Yorks.
Atlas, J. (1989), Philosophy without ambiguity. Clarendon Press, Oxford.
Austin,]. (1962), How to do things with words. Clarendon Press, Oxford.
Austin, J. P. M (1989), The dark side of politeness: A pragmatic analysis of nonco-operative communication. University of Canterbury, New Zealand, Ph.D.
thesis.
Bach, K. (1994a), 'Semantic slack: what is said and more', in Tsohatzidis 1994:
267-9l.
Bach, K. (1994b), 'Conversational implicature'. Mind and Language 9: 124-62.
Bach, K. and Harnish, R. (1979), Linguistic communication and speech acts.
MIT Press, Cambridge, MA.
Barkow, J., Cosmides, L. and Tooby, J. (1992), The adapted mind: Evolutionary
psychology and the generation of culture. Oxford University Press, New
York, N.Y.
de Beaugrande, R. and DressIer, W. (1981), Introduction to text linguistics.
Longman, London.
Bender, J. and Wellbery, D. (eds) (1990), The ends of rhetoric: History, theory,
practice. Stanford University Press, Stanford, CA.
300
Bibliography
Bennett, J. (1976), Linguistic behaviour. Cambridge University Press,

Cambridge.
Berg, J. (1991), 'The relevant relevance'.}ournal of Pragmatics 16.5: 411-25.
Bertolet, R. (1990), What is said. Kluwer, Dordrecht.
Bertuccelli-Papi, M. (1992), 'Determining the proposition expressed by an
utterance: The role of "domain adverbs"'. Textus V: 123-40.
Bever, T., Katz, J. and Langendoen, T. (eds) (1976), An integrated theory of
linguistic ability. Crowell, New York.
Bird, G. (1994), 'Relevance theory and speech acts', in Tsohatzidis 1994:
292-311.
Black, M. (ed.) (1965), Philosophy in America. Allen &: Unwin, London.
Blackburn, S. (ed.) (1975), Meaning, reference and necessity. Cambridge University Press, Cambridge.
Blackburn, S. (1984), Spreading the word. Oxford University Press, Oxford.
Blakemore, D. (1985), 'Discourse connectives and conjoined utterances'. Paper
presented to the Linguistics Association of Great Britain, September.
Blakemore, D. (1987), Semantic constraints on relevance. Blackwell, Oxford.
Blakemore, D. (1988a), '''So'' as a constraint on relevance', in Kempson 1988a:
183-95.
Blakemore, D. (1988b), 'The organization of discourse', in Newmeyer 1988,
vol. IV: 229-50.
Blakemore, D. (1989a), 'Linguistic form and pragmatic interpretation: The
explicit and the implicit', in Hickey 1989: 28-51.
Blakemore, D. (1989b), 'Denial and contrast: A relevance-theoretic analysis of
"but'''. Linguistics and Philosophy 12: 15-37.
Blakemore, D. (1990), 'Constraints on interpretation'. Proceedings of the 16th
annual meeting of the Berkeley Linguistic Society: General session and
parasession on the legacy of Grice: 363-70.
Blakemore, D. (1990/1), 'Performatives and parentheticals'. Proceedings of the
Aristotelian Society XCI.3: 197-213.
Blakemore, D. (1992), Understanding utterances: An introduction topragmatics.
Blackwell, Oxford.
Blakemore, D. (1993), 'The relevance of reformulations'. Language and Literature 2.2: 101-20.
Blakemore, D. (1994a), 'Relevance, poetic effects and social goals: A reply to
Culpeper'. Language and Literature 3.1: 49-59.
Blakemore, D. (1994b), 'Echo questions: A pragmatic account'. Lingua 94:
197-211.
BI~~emore, D. (forthcoming), 'Relevance theory'. To appear in Verschueren,
Ostman and Blommaert forthcoming.
Blass, R. (1986), 'Cohesion, coherence and relevance' . Notes on Linguistics 34:
41-64.
Blass, R. (1989), 'Grammaticalisation of interpretive use: The case of rE in
Sissala'. Lingua 79: 229-326.
Blass, R. (1990), Relevance relations in discourse: A study with special reference
to Sissala. Cambridge University Press, Cambridge.
Blass, R. (1993), 'Are there logical relations in a text?' Lingua 90.1/2: 91-110.
Bibliography
301
Booth, W. (1974), A rh,etoric of irony. C~icago University Press, Chicago.

Brockway, D. (1981), Semantic constramts on relevance', in Parret, Sbisa and
Verschueren 1981: 57-78.
Brockway, D. (1983), 'Pragmatic connectives'. Paper presented to the Linguistics Association of Great Britain, April.
Brown, G., Malkmjaer, K., Pollitt, A. and Williams, ]. (1994), Language and
understanding. Oxford University Press, Oxford.
Brown, G. and Yule, G. (1983), Discourse analysis. Cambridge University Press,
Cambridge.
Burkhardt, A. (ed.) (1990), Speech acts, meaning and intentions: Critical
approaches to the philosophy ofJohn Searle. Waiter de Gruyter, Berlin.
Burton-Roberts, N. (1985), 'Utterance, relevance and problems with text
grammar'. Australian Journal of Linguistics 5.2: 285-96.
Burton-Roberts, N. (1989a), The limits to debate: A revised theory of semantic
presupposition. Cambridge University Press, Cambridge.
Burton-Roberts, N. (1989b), 'On Horn's dilemma: Presupposition and
negation'. Journal of Linguistics 25: 95-125.
Cairns, H. and Kamerman, ]. (1975), 'Lexical information processing during
sentence comprehension'. Journal of Verbal Learning and Verbal Behavior
14: 170-9.
Campbell, J. (1990), 'The relevant communication of rhetorical arguments'.
Dissertation Abstracts International 51.6: 2001A.
Campbell, J. (1992), 'An applied relevance theory of the making and understanding of rhetorical arguments'. Language and Communication 12.2: 145-55.
Carnap, R. (1950), Logical foundations of probability. Routledge & Kegan Paul,
London.
Carston, R. (1984a), Review of Pulman 1983. Australian Journal of Linguistics
4.1: 89-99.
Carston, R. (1984b), 'Semantic and pragmatic analyses of "and" '. Paper
delivered to the Linguistics Association of Great Britain, April.
Carston, R. (1988a), 'Implicature, explicature and truth-theoretic semantics', in
Kempson 1988a: 155-81. Reprinted in Davis 1991: 33-5l.
Carston, R. (1988b), 'Language and cognition', in Newmeyer 1988, vol. Ill:
38-68.
Carston, R. (1993a), 'Syntax and pragmatics', in The encyclopedia of language
and linguistics. Pergamon Press and Aberdeen University Press, Oxford and
Aberdeen.
Carston, R. (1993b), 'Conjunction and pragmatic effects', in The encyclopedia
of language and linguistics. Pergamon Press and Aberdeen University Press,
Oxford and Aberdeen.
Cars ton, R. (1993c), 'Conjunction, explanation and relevance'. Lingua 90.112:
27-48.
Carston, R. (forthcoming a), 'Metalinguistic negation and echoic use'. To appear
in Journal of Pragmatics.
Carston, R. (forthcoming b), 'Quantity maxims and generalised implicature'.
To appear in Lingua.
302
Bibliography
Carston, R., Uchida, S. and Song, N. S. (forthcoming) Proceedings of the Osaka

workshop on relevance theory. John Benjamins, Amsterdam.
Chafe, W. (1976), 'Givenness, contrastiveness, definiteness, subjects, topics and
points of view', in Li 1976: 25-55.
Chametzky, R. (1992), 'Pragmatics, prediction and relevance'. Journal of
Pragmatics 17.1: 63-81.
Charolles, M. (1990), 'Cout, surcout et pertinence'. Cahiers de linguistique
franr;aise 11: 127-47.
Chiappe, D. and Kukla, A. (forthcoming), 'Context-selection and the frame
problem'. To appear in The Behavioral and Brain Sciences.
Clark, B. (1991), Relevance theory and the semantics of non-declaratives.
University of London, Ph.D. thesis.
Cl ark, B. (1993a), 'Relevance and "pseudo-imperatives"'. Linguistics and Philosophy 16: 79-12l.
Clark, B. (1993b), 'Let and let's: Procedural encoding and explicature'. Lingua
90.112: 173-200.
Clark, H. (1977), 'Bridging', in Johnson-Laird and Wason 1977: 411-20.
Clark, H. (1978), 'Inferring what is meant', in Levelt and Flores d' Arcais 1978:
295-322.
Clark, H. and Carlson, T. (1981), 'Context for comprehension', in Long and
Baddeley 1981: 313-30.
Clark, H. and Gerrig, R. (1984), 'On the pretense theory of irony'. Journal of
Experimental Psychology: General 113.1 : 121-6.
Clark, H. and Haviland, S. (1977), 'Comprehension and the given-new contract', in Freedle 1977: 1-40.
Cl ark, H., and Lucy, P. (1975), 'Understanding what is meant from what is
said: A study in conversationally conveyed requests'. Journal of Verbal
Learning and Verbal Behavior 14: 56-72.
Clark, H. and Marshall, C. (1981), 'Definite reference and mutual knowledge',
in Joshi, Webber and Sag 1981: 10-63.
Clark, H. and Schunk, D. (1980), 'Polite responses to polite requests'. Cognition
8.2: 111-43.
Cole, P. (ed.) (1978), Syntax and semantics 9: Pragmatics. Academic Press, New
York.
Cole, P. (ed.) (1981), Radical pragmatics. Academic Press, New York.
. Cole, P. and Morgan, J. (eds) (1975), Syntax and semantics 3: Speech acts.
Academic Press, N ew York.
Collinge, N. (ed.) (1990), An encyclopedia of language. Routledge, London.
Cormack, A. (1980), 'Negation, ambiguity and logical form'. University College
London.
Culpeper, J. (1994), 'Why relevance theory does not explain "the relevance of
reformulations'''. Language and Literature 3.1: 43-8.
Dancy, J., Moravcsik, J. and Taylor, C. (1988), Human agency: Language, duty
and value. Stanford University Press, Stanford, CA.
Davidson, D. (1984a), 'Communication and convention', in Davidson 1984b:
265-80.
Davidson, D. (1984b), Truth and interpretation. Clarendon Press, Oxford.
Bibliography
303
Davidson, D. and Harman, G. (eds) (1972), The semantics of natural language.

Reidel, Dordrecht.
Davies, M. (1981), Meaning, quantification, necessity: Themes in philosophical
logic. Routledge & Kegan Paul, London.
Davis, S. (ed.) (1991), Pragmatics: A reader. Oxford University Press, Oxford.
Demonte, V. and Garza Cuaron, B. (eds) (1990), Estudios de linguistica de
Espana y Mexico. Universidad Nacional Autonoma de Mexico, Mexico City.
Dretske, F. (1981), Knowledge and the flow of information. Blackwell, Oxford.
Ducrot, O. (1972), Dire et ne pas dire. Hermann, Paris.
Ducrot, O. (1980a), 'Analyses pragmatiques'. Communications 32: 11-60.
Ducrot, O. (1980b), Les echelles argumentatives. Minuit, Paris.
Ducrot, O. (1983), 'Puisque: essai de description polyphonique', in Herslund et
al.1983: 166-85.
Ducrot, O. (1984), Le dire et le dit. Minuit, Paris.
Durant, A. and Fabb, N. (1990), Literary studies in action. Routledge, Londo.n.
Escandell Vidal, M. V. (1993), Introduction a la pragmdtica. Editorial Anthropos, Barcelona.
Espinal, T. (1985), Analysas interpretives i teoria lingiiistica. University Autonoma, Barcelona, Ph.D. thesis.
Espinal, T. (1991), 'The representation of disjunct constituents', Language 67:
726-62.
Espinal, T. (1993), 'The interpretation of no pas in Catalan'. Journal of
Evans, G. and McDowell,]. (eds) (1976), Truth and meaning. Oxford University Press, Oxford.
Fabb, N. (forthcoming), Linguistics and literary theory. Blackwell, Oxford.
Fabb, N. (in preparation), Insight and arousal: The cognitive structure of intense
aesthetic experience.
Fabb, N., Attridge, D., Durant, A. and McCabe, C. (eds) (1987), The linguistics
of writing: Arguments between language and literature. Manchester Univer-~
sity Press, Manchester.
Ferrar, M. (1993), The logic of the ludicrous: A pragmatic study of humour.
Fodor,]. A. (1974), 'Special sciences'. Synthese 28: 77-115. Reprinted in Fodor
1981b: 127-45.
Fodor, J. A. (1975), The language of thought. Crowell, New York.
Fodor,]. A. (1980), 'Methodological solipsism considered as a research strategy
in cognitive psychology'. The Behavioral and Brain Sciences 3.1: 63-109.
Reprinted in Fodor 1981b: 225-53.
Fodor, ]. A. (1981a), 'The present status of the innateness controversy', m
Fodor 1981b: 257-316.
Fodor, J. A. (1981 b), Representations. Harvester Press, Hassocks.
Fodor, J. A. (1982), 'Cognitive science and the twin-earth problem'. Notre
Dame Journal of Formal Logic 23.2: 98-118.
Fodor, ] .A. (1983), The modularity of mind. MIT Press, Cambridge, MA.
Fodor, ]. A., Garrett, M., Walker, E. and Parkes, C. (1980), 'Against definitions'. Cognition 8.3: 263-367.
304
Bibliography
Fodor, J. D. (1977), Semantics: Theories of meaning in generative grammar.

Harvester Press, Hassocks.
Fodor, J. D. (1979), 'In defense of the truth-value gap', in Oh and Dinneen
1979: 199-224.
Forceville, C. (1994a), Pictorial metaphor in advertising. Vrije Universiteit,
Amsterdam.
Forceville, C. (1994b), 'Pictorial metaphor in billboards: Relevance theory
perspectives', in Muller 1994: 93-113.
Forget, D. (1989), 'La: un marqueur de pertinence discursive'. Revue quebecoise
de linguistique 18.1: 57-83.
Foster, M. and Brandes, S. (eds) (1980), Symbol as sense. Academic Press, New
York.
Foster-Cohen, S. (1994), 'Exploring the boundary between syntax and pragmatics: Relevance and the binding of pronouns'. Journal of Child Language 21:
237-55.
Fowler, A. (1989), 'Review of Relevance: Communication and cognition'.
London Review of Books, 30 March.
Freedle, R. (ed.) (1977), Discourse production and comprehension. Ablex,
Norwood, NJ.
Fretheim, T. (forthcoming a), 'Accessing contexts with intonation'. To appear
in Fretheim and Gundel (forthcoming).
Fretheim, T. (forthcoming b), 'Pragmatic implications of "not until" in Norwegian'. To appear in Simonsen, Loedrup and Moen (forthcoming).
Fretheim, T. and Gundel, J. (eds) (forthcoming), Reference and referent
accessibility. John Benjamins, Amsterdam.
von Frisch, K. (1967), The dance language and orientation of bees. Belknap
Press of Harvard University Press, Cambridge, MA.
Frith, U. (1989), Autism: Explaining the enigma. Blackwell, Oxford.
Frith, U. (ed.) (1991), Autism and Asperger syndrome. Cambridge University
Press, Cambridge.
Gabbay, D. and Kempson, R. (1991), 'Labelled abduction and relevance
reasoning', in SOAS Working Papers in Linguistics and Phonetics, 2: 41-84.
To appear in Proceedings of the workshop on non-standard queries and
answers (Toulouse, 1991).
Gardiner, A. H. (1932), The theory of speech and language. Oxford University
Press, Oxford.
Garnham, A. and Perner, J. (1990), 'Does manifestness solve problems of
mutuality?' The Behavioral and Brain Sciences 13.1: 178-9.
Gazdar, G. (1979), Pragmatics: Implicature, presupposition and logical form.
Academic Press, New York.
Gazdar, G. and Good, D. (1982), 'On a notion of relevance', in Smith 1982:
88-100.
Gibbs, R. (1987), 'Mutual knowledge and the psychology of conversational
inference'. Journal of Pragmatics 11.5: 561-88.
Gibbs, R. (1994), The poetics of mind: Figurative thought, language and
understanding. Cambridge University Press, Cambridge.
Bibliography
305
Giora, R. (1988), 'On the informativeness requirement'. Journal of Pragmatics

12.5/6: 547-65.
Givon, T. (ed.) (1979), Syntax and semantics 12: Discourse and syntax. Academic
Press, New York.
Goatly, A. (1994), 'Register and the redemption of relevance theory: The case
of metaphor'. Journal of the International Pragmatics Association 4.2: 139-8l.
Goodman, N. (1955), Fact, fiction and forecast. Harvard University Press,
Cambridge, MA.
Gorayska, B. and Lindsey, R. (1993), 'The roots of relevance'. Journal of
Gordon, D. and Lakoff, G. (1975), 'Conversational postulates', in Cole and
Morgan 1975: 83-106.
Green, G. (1980), 'Some wherefores of English inversions'. Language 56:
582-60l.
Green, G. and Morgan, J. (1981), 'Pragmatics, grammar and discourse', in Cole
1981: 167-8l.
Green, K. (1993), 'Relevance theory and the literary text: Some problems and
perspectives.' Journal of Literary Semantics 22: 207-17.
Grice, H. P. (1957), 'Meaning'. Philosophical Review 66: 377-88. Reprinted in
Steinberg and Jakobovits 1971: 53-9 and Grice 1989: 213-23.
Grice, H. P. (1961), 'The causal theory of perception'. Proceedings of the
Aristotelian Society, Supplementary vol. 35: 121-52. Reprinted in Grice 1989:
224-47.
Grice, H. P. (1967), Logic and conversation. William James Lectures, reprinted
in Grice 1989: 1-143.
Grice, H. P. (1968), 'Utterer's meaning, sentence meaning and word meaning'.
Foundations of Language 4: 225-42. Reprinted in Searle 1971: 54-70 and
Grice 1989: 117-37.
Grice, H. P. (1969), 'Utterer's meaning and intentions'. Philosophical Review
78: 147-77. Reprinted in Grice 1989: 86-116.
Grice, H. P. (1975), 'Logic and conversation', in Cole and Morgan 1975: 41-58.
Reprinted in Grice 1989: 22-40.
Grice, H. P. (1978), 'Further notes on logic and conversation', in Cole 1978:
113-28. Reprinted in Grice 1989: 41-57.
Grice, H. P. (1981), 'Presupposition and conversational implicature', in Cole
1981: 183-98. Reprinted in Grice 1989: 269-82.
Grice, H. P. (1982), 'Meaning revisited', in Smith 1982: 223-43. Reprinted in
Grice 1989: 283-303.
Grice, H. P. (1989), 'Retrospective epilogue', in Grice 1989: 339-85.
Grice, H. P. (1989), Studies in the way of words. Harvard University Press,
Cambridge, MA.
Grimshaw, J. (1979), 'Complement selection and the lexicon'. Linguistic Inquiry
10.2: 279-326.
Groefsema, M. (1992a), Processing for relevance: A pragmatically based account
of how we process natural language. University of London, Ph.D. thesis.
Groefsema, M. (1992b), "'Can you pass the salt?": A short-circuited implicature?' Lingua 87: 103-35.
306
Bibliography
Groefsema, M. (1995), 'Can, may, must and should: A relevance-theoretic

approach'. Journal of Linguistics 31: 53-79.
Groefsema, M. (forthcoming), 'Understood arguments: A semantic-pragmatic
approach. To appear in Lingua.
Grundy, P. (1995), Doing pragmatics. Edward Arnold, London.
Guijarro Morales, J.-L. (ed.) (1993), Pragmalingiiistica 1. University of Cadiz,
Cadiz.
Gundel, J. (forthcoming), 'Relevance theory meets the givenness hierarchy: An
account of inferrables'. To appear in Fretheim and Gundel forthcoming.
Gunderson, K. (ed.) (1975), Language, mind and knowledge. Minnesota Studies
in the Philosophy of Science, VII. University of Minnesota Press, Minneapolis, MN.
Gussenhoven, C. (1983), 'Focus, mode and the nucleus'. Journal of Linguistics
19.2: 377-417.
Gutt, E.-A. (1986), 'Unravelling meaning: An introduction to relevance theory'.
Notes on Translation 112: 10-20.
Gutt, E.-A. (1988), 'Towards an analysis of pragmatic connectives in Silt'i'.
Proceedings of the Eighth International Conference of Ethiopian Studies.
Addis Ababa University: 26-30.
Gutt, E.-A. (1990), 'A theoretical account of translation: Without a translation
theory'. Target 2.2: 135-64.
Gutt, E.-A. (1991), Translation and relevance: Cognition and context. Blackwell, Oxford.
Gutt, E.-A. (1992), Relevance theory: A guide to successful communication in
translation. Lectures delivered at the Triennial Translation Workshop of UBS,
Zimbabwe. Summer Institute of Linguistics, Dallas; United Bible Societies,
New York.
Haegeman, L. (1987), 'The interpretation of inherent objects in English'.
Australian Journal of Linguistics 7.2: 223-48.
Haegeman, L. (1989), 'Be going to and will: A pragmatic account'. Journal of
Linguistics 25.2: 291-319.
Haegeman, L. (1993), 'The interpretation of the particle da in West Flemish'.
Lingua 90.1/2: 111-28.
Halle, M., Bresnan, J. and Miller, G. (eds) (1978), Linguistic theory and
psychological reality. MIT Press, Cambridge, MA.
Halliday, M. (1967/8), 'Notes on transitivity and theme in English'. Journal of
Linguistics 3: 37-81; 4: 179-215.
Hamamoto, H. (forthcoming), 'Irony from a cognitive perspective'. To appear
in Carston, Uchida and Song forthcoming.
Happe, F. (1991), 'The autobiographical writings of three Asperger syndrome
adults: Problems of interpretation and implications for theory', in Frith 1991:
207-42.
Happe, F. (1992), Theory of mind and communication in autism. University of
London, Ph.D. thesis.
Happe, F. (1993), 'Communicative competence and theory of mind in autism:
A test of relevance theory'. Cognition 48.2: 101-19.
Bibliography
307
Harman, G. (19?8), '"!,hree.levels of meaning'. Journal of Philosophy LXV:

590-602. Repnnted m Stemberg and Jakobovits 1971: 66-75.
Harnish, R. M. (1976), 'Logical form and implicature', in Bever, Katz and
Langendoen 1976: 464-79. Reprinted in Davis 1991: 316-64.
Harnish, R. M. (1994), 'Mood, meaning and speech acts', in Tsohatzidis 1994:
407-59.
Hawkins, J.
(1991), 'On (in)definite articles: Implicatures and
(un)grammaticality predictions.' Journal of Linguistics 27: 405-42.
Herslund, M. et al. (eds) (1983), Analyses grammaticales du franqais. Special
issue of Revue romane, 24.
Hickey, L. (ed.) (1989), The pragmatics of style. Routledge, London.
Higashimori, I. (1992a), 'Review of Kempson (ed.) M ental Representations'.
English Linguistics 9: 335-56.
Higashimori, I. (1992b), 'Except, but and relevance theory'. English Literature
Review (Kyoto Women's University) 36: 62-108.
Higashimori, I. (1994), 'A relevance-theoretic analysis of even, saelsuralmol
temolddemoldattelmade'. English Literature Review (Kyoto Women's University) 38: 51-80.
Hirschfeld, L. and Gelman, S. (1994), Mapping the mind: Domain specificity in
cognition and culture. Cambridge University Press, Cambridge.
Hirst, D. (1987), 'Intonation: Syntaxe, semantique et pragmatique'. Sigma 11:
148-70.
Hirst, D. (1989), 'Review of Relevance.' Mind and Language 4.112: 138-46.
Hjelmslev, L. (1928), Principes de grammaire generale. Akademisk Foriag,
Copenhagen.
Hjelmslev, L. (1959), Essais linguistiques. Akademisk Foriag, Copenhagen.
Hobbs, J. (1979), 'Coherence and coreference'. Cognitive Science 3.1: 67-90.
Hogaboam, T. and Perfetti, C. (1975), 'Lexical ambiguity and sentence comprehension'.Journal of Verbal Learning and Verbal Behavior 14: 265-74.
Horn, L. (1984), 'A new taxonomy for pragmatic inference: Q-based and Rbased implicature', in Schiffrin 1984: 11-42.
Horn, L. (1988), 'Pragmatic theory', in Newmeyer 1988, vo!. I: 113-45.
Horn, L. (1992), 'The said and the unsaid'. Ohio State University Working
Papers in Linguistics 40: 163-92.
Hugly, P. and Sayward, C. (1979), 'A problem about conversational implicature'. Linguistics and Philosophy 3: 19-25.
Hyman, L. and Li, C. (1988), Language, speech and mind: Essays in honor of
Victoria Fromkin. Routledge, London.
Hymes, D. (1987), 'A theory of verbal irony and a Chinookan pattern of verbal
exchange', in Verschueren and Bertuccelli-Papi 1987: 293-337.
Ibsch, E., Schram, D. and Steen, G. (eds) (1991), Empirical studies of literature:
Proceedings of the second fGEL conference (Amsterdam 1989). Rodopi,
Amsterdam.
Ifantidou, E. (1994), Evidentials and relevance. University of London, Ph.D.
thesis.
Ifantidou-Trouki, E. (1993), 'Sentential adverbs and relevance'. Lingua 90.112:
65-90.
308
Bibliography
Itani, R. (1993), 'The Japanese sentence-final particle ka: A relevance-theoretic

approach'. Lingua 90.112: 129-47.
Itani, R. (1995), Semantics and pragmatics of hedges in English and Japanese.
Jackendoff, R. (1972), Semantic interpretation in generative grammar. MIT
Press, Cambridge, MA.
Jakobson, R. (1960), 'Linguistics and poetics', in Sebeok 1960: 350-77.
Jayez, J. (1986), 'L'analyse de la notion de pertinence d'apres Sperber et Wilson'.
Sigma 10: 7-46.
Jiang, y. (1994), 'A procedural account of Chinese quantification'. Paper
presented to the Third International Conference on <Shinese Linguistics, City
Polytechnic of Hong Kong. To appear in Selected Papers of ICCL-3.
Jodlowiec, M. (1991), The role of relevance in the interpretation of verbal jokes:
a pragmatic analysis, Jagiellonian University, Krakow, Ph.D. thesis.
Johnson-Laird, P. (1967), An experimental investigation into one pragmatic
factor governing the use of the English language. University of London,
Ph.D. thesis.
Johnson-Laird, P. (1982a), 'Mutual ignorance: Comments on Clark and Carlson's paper', in Smith 1982: 40-5.
Johnson-Laird, P. (1982b), 'Thinking as a skill'. Quarterly Journal of Experimental Psychology 34A: 1-29.
Johnson-Laird, P. (1983), Mental models. Cambridge University Press,
Cambridge.
Johnson-Laird, P. and Wason, P. (eds) (1977), Thinking: Readings in cognitive
science. Cambridge University Press, Cambridge.
Jorgensen, J., Miller, G. and Sperber, D. (1984), 'Test of the mention theory of
irony'. Journal of Experimental Psychology: General 113.1 : 112-20.
Joshi, A., Webber, B. and Sag, I. (eds) (1981), Elements of discourse understanding. Cambridge University Press, Cambridge.
Jucker, A. (1988), 'The relevance of politeness'. Multilingua 7.4: 375-84.
Jucker, A. (1993), 'The discourse marker well: A relevance-theoretical account'.
Journal of Pragmatics 19: 435-52.
Just, M. and Carpenter, P. (eds) (1977), Cognitive processes in comprehension.
Lawrence Erlbaum, Hillsdale, NJ.
Kahneman, D., Slovic, P. and Tversky, A. (1982), Judgement under uncertainty:
Heuristics and biases. Cambridge University Press, Cambridge.
Kandolf, C. (1993), 'On the difference between explicatures and implicatures in
relevance theory'. NordicJournal of Linguistics 16.1: 33-46.
Kasher, A. (1994), 'Modular speech-act theory: Programme and results', in
Tsohatzidis 1994: 312-22.
Kasher, A. (ed.) (forthcoming), Pragmatics: Critical assessment. Routledge,
London.
Katz, J. (1972), Semantic theory. Harper & Row, New York.
Katz, J. (1977), Propositional structure and illocutionary force: A study of the
contribution of sentence meaning to speech acts. Harvester Press, Hassocks.
Katz, J. (1981), Language and other abstract objects. Blackwell, Oxford.
Bibliography
309
Kempson, R. (1975), Presupposition and the delimitation of semantics. Cambridge University Press, Cambridge.
Kempson, R. (1984), 'Anaphora, the compositionality requirement and the
semantics-pragmatics distinction'. Proceedings of the North-Eastern Linguistics Society XIV, University of Massachusetts, Amherst, MA.
Kempson, R. (1986), 'Ambiguity and the semantics-pragmatics distinction', in
Travis 1986: 77-103.
Kempson, R. (ed.) (1988a), Mental representations: The interface between
language and reality. Cambridge University Press, Cambridge.
Kempson, R. (1988b), 'Grammar and conversational principles', in Newmeyer
1988, vol. 11: 139-63.
Kempson, R. (1988c), 'Logical form: The grammar-cognition interface'. Journal
of Linguistics 24.2: 393-431.
Kempson, R. (forthcoming), 'Semantics, pragmatics and natural-language interpretation' . To appear in Lappin forthcoming.
Kerbrat-Orecchioni, C. (1977), La connotation. Presses Universitaires de Lyon,
Lyon.
Kerbrat-Orecchioni, C. (1981), 'L'ironie comme trope'. Poetique 41: 108-27.
Khalfa, J. (ed.) (1994), What is intelligence? Cambridge University Press,
Cambridge.
Kiparsky, P. (1987), 'On theory and interpretation', in Fabb, Attridge, Durant
and McCabe 1987: 185-98.
Kirschner, C. and de Cesaris, J. (eds) (1989), Studies in Romance linguistics.
John Benjamins, Amsterdam.
Kleiber, G. (1990), 'Marqueurs referentiels et processus interpretatifs: pour une
approche "plus semantique'''. Cahiers de linguistique franr;aise 11: 241-58.
Kleiber, G. (1992), 'Article defini: unicite et pertinence'. Revue roumane 27.1:
61-89.
Klinge, A. (1993), 'The English modal auxiliaries: From lexical semantics to
utterance interpretation'. Journal of Linguistics 29: 315-57.
Konig, E. (1991a), 'A relevance-theoretic approach to the analysis of modal
particles in German'. Multilingua 10. 112: 63-77.
Konig, E. (1991b), The meaning of focus particles: A comparative perspective.
Routledge, London.
Kreuz, R. and Glucksberg, S. (1989), 'How to be sarcastic: The echoic reminder
theory of verbal irony' . Journal of Experimental Psychology: General 118:
374-86.
Kripke, S. (1972), 'Naming and necessity', in Davidson and Harman 1972:
253-355. Reprinted as Naming and necessity. Blackwell, Oxford (1980).
Lappin, S. (ed.) (forthcoming), Handbook of contemporary semantics. Blackwell,
Oxford.
Leach, E. (1976), Culture and communication. Cambridge University Press,
Cambridge.
Leech, G. (1983), Principles of pragmatics. Longman, London.
Leech, G. and Thomas, J. (1990), 'Language, meaning and context: Pragmatics',
in Collinge 1990: 173-205.
Lemmon, E. (1965), Beginning logic. Nelson, London.
310
Bibliography
Leslie, A. (1989), 'Review of Relevance'. Mind and Language 4. 112: 147-50.

Levelt, W. and Flores d'Arcais, G. (eds) (1978), Studies in the perception of
language. John Wiley, Chichester.
Levinson, S. (1983), Pragmatics. Cambridge University Press, Cambridge.
Levinson, S. (1987), 'Minimization and conversational inference', in Verschueren and Bertuccelli-Papi 1987: 61-129.
Levinson, S. (1988), 'Generalised conversational implicatures and the
semantics-pragmatics interface'. Cambridge.
Levinson, S. (1989), 'A review of Relevance'. Journal of Linguistics 25.2: 455-72.
Lewis, D. (1969), Convention. Harvard University Press, Cambridge, MA.
Lewis, D. (1975), 'Languages and language', in Gunderson 1975: 3-35.
Reprinted in Lewis 1983: 163-88.
Lewis, D. (1983), Philosophical papers, vol. 1. Oxford University Press, Oxford.
Li, C. (ed.) (1976), Subject and topic. Academic Press, New York.
Loar, B. (1976), 'Two theories of meaning', in Evans and McDowell 1976:
138-6l.
Loar, B. (1981), Mind and meaning. Cambridge University Press, Cambridge.
Loftus, E. (1979), Eyewitness testimony. Harvard University Press, Cambridge,
MA.
Long, J. and Baddeley, A. (eds) (1981), Attention and performance IX. Lawrence
Erlbaum, Hillsdale, NJ.
Lunn, P. (1989), 'The Spanish subjunctive and relevance', in Kirschner and de
Cesaris 1989: 249-60.
Luscher, J.-M. (1994), 'Les marques de connexion: des guides pour l'interpretation', in Moeschler et al. 1994: 175-227.
Lyons, J. (1977), Semantics. Cambridge University Press, Cambridge.
MacLaran, R. (1982), The semantics and pragmatics of English demonstratives.
Cornell University, Ph.D. thesis.
Mao, L. (1992), Pragmatic universals and their implications. University of
Minnesota, Ph.D. thesis. Dissertation Abstracts International 52.8: 2908A.
Marslen-Wilson, W. (1973), Speech shadowing and speech perception. MIT,
Ph.D. thesis.
Marslen-Wilson, W. and Tyler, L. (1980), 'The temporal structure of spoken
language understanding'. Cognition 8.1: 1-72.
Martin, R. (1992), 'Irony and universe of belief'. Lingua 87: 77-90.
Matsui, T. (1993), 'Bridging reference and the notions of topic and focus'.
Lingua 90.112: 49-68.
Matsui, T.(1995), Bridging and relevance. University of London, Ph.D. thesis.
Mayher, J. (1990), Uncommon sense: Theoretical practice in language education.
Heinemann, London.
McCawley, J. (1979), 'Presupposition and discourse structure', in Oh and
Dinneen 1979: 371-88.
McCawley, J. (1980), Everything that linguists have always wanted to know
about logic but were ashamed to ask. University of Chicago Press, Chicago,
IL.
McDowell, J. (1980), 'Meaning, communication and knowledge', in van Straaten
1980: 117-39.
Bibliography
311
Mey, J. (ed.) (1986), Language and discourse: Text and protest. John Benjamins,
Amsterdam.
Mey, J. and Talbot, M. (1988), 'Computation and the soul'. Journal of
Pragmatics 12: 743-89.
Miller, G. and Johnson-Laird, P. (1976), Language and perception. Cambridge
University Press, Cambridge.
Minsky, M. (1977), 'Frame system theory', in Johnson-Laird and Wason 1977:
355-76.
Mitsunobu, M. (1993), 'A relevance theory bibliography'. Department of
English, Tokyo Metropolitan University.
Moeschler, J. (1989a), 'Pragmatic connectives, argumentative coherence and
relevance'. Argumentation 3.3: 321-39.
Moeschler, J. (1989b), Modelisation du dialogue. Representation de l'inference
argumentative. Hermes, Paris.
Moeschler, J. (ed.) (1989c), Argumentation, relevance and discourse. Argumentation 3.3. Kluwer, Dordrecht.
Moeschler, J. (1991), 'The pragmatic aspects of linguistic negation: Speech acts,
argumentation and pragmatic inference'. Argumentation 6: 51-75.
Moeschler, J. (1992), 'Une, deux ou trois negations?' Langue franr;;aise 94: 8-25.
Moeschler, J. (1993a), 'Relevance and conversation'. Lingua 90. 112: 149-71.
Moeschler, J. (1993b), 'Aspects pragmatiques de la reference temporelle: indetermination, ordre temporel et inference'. Langages 112: 39-54.
Moeschler, J. and Reboul, A. (1994), Dictionnaire encyclopedique de pragmatique. Seuil, Paris.
Moeschler, J., Reboul, A., Luscher, J.-M. and Jayez, J. (1994), Langage et
pertinence: Reference temporelle, anaphore, connecteurs et metaphore. Presses
Universitaires de Nancy, Nancy.
Morgan, J. (1979), 'Observations on the pragmatics of metaphor', in Ortony
1979: 136-47.
Morris, C. (1938), 'Foundations of the theory of signs', in Neurath, Carnap and
Morris 1938: 77-138. Reprinted in Morris 1971.
Morris, C. (1971), Writings on the general theory of signs. Mouton, The Hague.
Muller, J. (ed.) (1994), Towards a pragmatics of the audiovisual. NODUS,
Munster.
Munitz, M. and Unger, P. (eds) (1974), Semantics and philosophy: Studies in
contemporary philosophy. New York University Press, New York.
Nasta, D. (1991), Meaning in film: Relevant structures in soundtrack and
narrative. Peter Lang, Bern.
N eale, S. (1992), 'Paul Grice and the philosophy of language'. Linguistics and
Philosophy 15.5: 509-59.
Nebeska, I. (1991), 'Muze byt relevance postacujicim principem komunikace?'
Slovo a Slovesnost 52.2: 104-8.
Neisser, U. (ed.) (1982), Memory observed: Remembering in natural contexts.
W. H. Freeman, San Francisco, CA.
Neurath, 0., Carnap, R. and Morris, C. (eds) (1938), International encyclopaedia of unified science. University of Chicago Press, Chicago, IL.
312
Bibliography
Newmeyer, F. (1988), Linguistics: The Cambridge Survey, vols I-IV. Cambridge University Press, Cambridge.
Nishiyama, Y. (1992), 'Hatsuwa-kaishaku to Ninchi: on Kanrensei-riron nitsuite (Utterance interpretation and cognition: On Relevance Theory)', in
Anzai et al. 1992: 466-76.
Nishiyama, Y. (1993), 'Kontekusuto-kouka to Kanrensei: Kanrensei-riron no
Mondai-ten (Contextual effects and relevance: Some problems of relevance
theory)'. Eigo Seinen 139.5: 14-16.
Nishiyama, Y. (1995), 'Gengai no Imi wo Toraeru (How to read between the
lines)'. Gengo 24.4: 30-9.
N0lke, H. (1990), 'Pertinence et modalisateurs d'nonciation'. Cahiers de
linguistique fram;aise 11.
Oh, c.- K. and Dinneen, D. (eds) (1979), Syntax and semantics 11: Presupposition. Academic Press, New York.
O'Neill, J. (1988), 'Relevance and pragmatic inference'. Theoretical Linguistics
15: 241-61.
Ortony, A. (ed.) (1979), Metaphor and thought. Cambridge University Press,
Cambridge.
Parret, H. (ed.) (1994), Pretending to communicate. WaIter de Gruyter, Berlin.
Parret, H., Sbisa, M. and Verschueren, J. (1981), Possibilities and limitations of
pragmatics. John Benjamins, Amsterdam.
Pateman, T. (1986), 'Relevance, contextual effects and least effort'. Poetics
Today 7.4: 745-54.
Patton, T. and Stampe, D. (1969), 'The rudiments of meaning: Ziff on Grice'.
Foundations of Language 5.1: 2-16.
Perrin, L. (forthcoming), L'ironie mise en trope, du sens litteral au sens figure.
University of Geneva, doctoral dissertation, to be published by Editions
Kime.
Philippaki-Warburton, I., Nicolaides, K. and Sifianou, M. (eds) (1994), Themes
in Greek linguistics. Papers from the 1st international conference on Greek
Linguistics (Reading 1993). John Benjamins, Amsterdam.
Pilkington, A. (1991), 'The literary reading process: A relevance theory perspective', in Ibsch, Schram and Steen 1991: 117-23.
Pilkington, A. (1992), 'Poetic effects'. Lingua 87. 112: 29-51.
Pilkington, A. (1994), Poetic thoughts and poetic effects. University of London,
Ph.D. thesis.
Pike, K. (1967), Language in relation to a unified theory of the structure of
human behavior. Mouton, The Hague.
Politzer, G. (1990), 'Characterizing spontaneous inferences'. The Behavioral
and Brain Sciences 13.1: 177-8.
Politzer, G. (1993), La psychologie du raisonnement: Lois de la pragmatique et
de la logique formelle. These de Doctorat d'Etat, University of Paris VIII.
Posnanski, V. (1992), A relevance-based utterance processing system. Cambridge
University, Ph.D. thesis, University of Cambridge Laboratory, Technical
report, No. 246.
Prince, E. (1981), 'Towards a taxonomy of given-new information', in Cole
1981: 223-56.
Bibliography
313
Pulman, S. (1983), Word meaning and belief Croom Helm, London.
Putnam, H. (1975a), 'The meaning of "meaning''', in Gunderson 1975: 131-93.
Reprinted in Putnam 1975b: 215-7l.
Putnam, H. (1975b), Mind, language and reality: Philosophical papers II.
Cambridge University Press, Cambridge.
'
Quine, W. (1960), Word and object. MIT Press, Cambridge, MA.
Reboul, A. (1990a), Analyse de la metaphore et de la fiction. University of
Geneva, Ph.D. thesis.
Reboul, A. (1990b), 'The logical status of fictional discourse: What Searle's
speaker can't say to his hearer', in Burkhardt 1990: 336-63.
Reboul, A. (1992a), Rhetorique et stylistique de la fiction. Presses Universitaires
de Nancy, Nancy.
Reboul, A. (1992b), 'How much am I I and how much is she I?' Lingua 87:
169-202.
Reboul, A. (1994a), 'L'anaphore pronominale: le probleme de l'attribution des
referents', in Moeschler et al. 1994: 105-73.
Reboul, A. (1994b), 'The description of lies in speech-act theory', in Parret
1994: 292-8.
Reboul, A. (forthcoming), 'What (if anything) is accessibility? A relevanceoriented criticism of Ariel's Accessibility Theory of referring expressions'. To
appear in Acts of the 6th international conference on functional grammar
(York, 1994).
Recanati, F. (1979), La transparence et l'enonciation. Seuil, Paris.
Recanati, F. (1987), Meaning and force. Cambridge University Press,
Cambridge.
Recanati, F. (1989), 'The pragmatics of what is said'. Mind and Language 4.4:
295-329. Reprinted in Davis 1991: 97-120.
Recanati, F. (1993), Direct reference: From language to thought. Blackwell,
Oxford.
Recanati, F. (1994), 'Contextualism and anti-contextualism in the philosophy of
language', in Tsohatzidis 1994: 156-66.
Recanati, F. (forthcoming), 'The alleged priority of literal interpretation'. To
appear in Cognitive Science.
Reddy, M. (1979), 'The conduit metaphor - a case of frame conflict in our
language about language', in Ortony 1979: 284-324.
Reinhart, T. (1981), 'Pragmatics and linguistics: An analysis of sentence topics'.
Philosophica 27: 53-94.
Richards, C. (1985), 'Inferential pragmatics and the literary text'. Journal of
Pragmatics 9: 261-85.
Rips, L. (1983), 'Cognitive processes in propositional reasoning'. Psychological
Review 90.1: 38-7l.
Roberts, L. (1991), 'Relevance as an explanation of communication'. Linguistics
and Philosophy 14.4: 453-72.
Rochemont, M. (1986), Focus in generative grammar. John Benjamins,
Amsterdam.
Rouchota, V. (1992), 'On the referential-attributive distinction'. Lingua 87.112:
137-67.
314
Bibliography
Rouchota, v. (1994a), The semantics and pragmatics of the subjunctive in

Modern Greek - a relevance-theoretic approach. University of London, Ph.D.
thesis.
Rouchota, V. (1994b), 'Na-interrogatives in Modern Greek: Their interpretation
and relevance.' In Philippaki-Warburton, I., Nicolaides, K. and Sifianou, M.
1994: 177-84.
Rouchota, V. (1994c), 'The subjunctive in Modern Greek: Dividing the labour
between semantics and pragmatics'. Journal of Modern Greek Studies 12:
185-20l.
Rouchota, V. (1994d), 'On indefinite descriptions'. Journal of Linguistics 30:
441-75.
Sadock, J. (1978), 'On testing for conversational implicature', in Cole 1978:
281-98.
Sadock, J. (1979), 'Figurative speech and linguistics', in Ortony 1979: 46-63.
Sadock, J. (1986), 'Remarks on the paper by Deirdre Wilson and Dan Sperber',
in Parasession on pragmatics and grammatical theory. Chicago Linguistics
Society 22: 85-90.
Sadock, J. and Zwicky, A. (1985), 'Speech-act distinctions in syntax', in Shop en
1985: 155-96.
Sag, I. (1981), 'Formal semantics and extra linguistic context', in Cole 1981:
273-94.
Sanchez de Zavala, V. (1990), 'Sobre la nueva teoria de la pertinencia', in
Demonte and Garza Cuaron 1990: 273-99.
Saussure, F. de (1974), Course in general linguistics, translated from the French
(1916) by Wade Baskin. Peter Owen, London.
Scancarelli, J. (1986), 'Interpretation in context: A cause of semantic change'.
Cahiers de l'Institut linguistique de Louvain 12.112: 167-82.
Schank, R. and Abelson, R. (1977), 'Scripts, plans and knowledge', in JohnsonLaird and Wason 1977: 421-32.
Schiffer, S. (1972), Meaning. Clarendon Press, Oxford.
Schiffrin, D. (ed.) (1984), Meaning, form and use in context. Georgetown
University Press, Washington, DC.
Searle, J. (1965), 'What is a speech act?', in Black 1965: 221-39.
Searle, J. (1969), Speech acts. Cambridge University Press, Cambridge.
Searle,J. (1971a), 'Introduction' to Searle 1971b: 1-12.
Searle,J. (ed.) (1971b), The philosophy of language. Oxford University Press,
Oxford.
Searle, J. (1975), 'Indirect speech acts', in Cole and Morgan 1975: 59-82.
Searle, J. (1979a), 'The classification of illocutionary acts', in Searle 1979b: 1-29.
Searle, J. (1979b), Expression and meaning. Cambridge University Press,
Cambridge.
Searle, J. (1983), Intentionality. Cambridge University Press, Cambridge.
Searle, J., Kiefer, F. and Bierwisch, M. (eds) (1980), Speech-act theory and
pragmatics. Reidel, Dordrecht.
Sebeok, T. (ed.) (1960), Style in language. MIT Press, Cambridge, MA.
Seto, K.-I. (forthcoming), 'On non-echoic irony'. To appear in Carston, Uchida
and Song forthcoming.
Bibliography
315
Seuren, P. (1987), 'The self-styling of relevance theory'. Journal of Semantics

5.2: 123-43.
Shannon, C. and Weaver, W. (1949), The mathematical theory of communication. University of Illinois Press, Urbana, IL.
Shop en, T. (ed.) (1985), Language typology and syntactic description. Cambridge
University Press, Cambridge.
Simonsen, H., Loedrup, H. and Moen, H. (forthcoming), Selected papers from
the XVth Scandinavian Conference of Linguistics (Oslo, 1995.)
Sinclair, M. (1995), 'Fitting pragmatics into the mind: Some issues in mentalist
pragmatics'. Journal of Pragmatics 23: 509-39.
Sin clair, M. and Winckler, W. (1991), Relevance theory: Explaining verbal
communication. Stellenbosch Papers in Linguistics, 18.
Smith, N. (ed.) (1982), Mutual knowledge. Academic Press, London.
Smith, N. (1983), 'On interpreting conditionals'. Australian Journal of Linguistics 3.1: 1-23.
Smith, N. (1989), The twitter machine. Blackwell, Oxford.
Smith, N. (1993), 'Observations sur la pragmatique des temps'. Langages 112:
26-38.
Smith, N. and Smith, A. (1988), 'A relevance-theoretic account of conditionals',
in Hyman and Li 1988: 322-52.
Smith, N. and Tsimpli, I. (1995), The mind of a savant. Blackwell, Oxford.
Smith, N. and Wilson, D. (1979), Modern linguistics: The results of Chomsky's
revolution. Penguin, Harmondsworth and Indiana University Press.
Smith, N. and Wilson, D. (1992), 'Introduction to the special issue on relevance
theory'. Lingua 87.112: 1-10.
Soames, S. (1979), 'A projection problem for speaker presuppositions'. Linguis.
tic Inquiry 10.4: 623-66.
Song, Nam-Sun (forthcoming), 'Metonymy and metaphor'. To appear in
Cars ton, U chida and Song forthcoming.
Sperber, D. (1975a), Rethinking symbolism. Cambridge University Press,
Cambridge.
Sperber, D. (1975b), 'Rudiments de rhetorique cognitive'. Poetique 23: 389-415.
Sperber, D. (1980), 'Is symbolic thought prerational?' in Foster and Brandes
1980: 25-44.
Sperber, D. (1984), 'Verbal irony: Pretense or echoic mention?' Journal of
Experimental Psychology: General 113.1 : 130-6.
Sperber, D. (1985), On anthropological knowledge. Cambridge University Press,
Cambridge.
Sperber, D. (1994a), 'Understanding verbal understanding', in Khalfa 1994:
179-98.
Sperber, D. (1994b), 'The modularity of thought and the epidemiology of
representations', in Hirschfeld and Gelman 1994: 39-67.
Sperber, D., Cara, F. and Girotto, V. (forthcoming), 'Relevance theory explains
the Selection Task'. To appear in Cognition.
Sperber, D. and Wilson, D. (1981), 'Irony and the use-mention distinction', in
Cole 1981: 295-318. Reprinted in Davis 1991: 550-63.
316
Bibliography
Sperber, D. and Wilson, D. (1982), 'Mutual knowledge and relevance in theories

of comprehension', in Smith 1982: 61-131.
Sperber, D. and Wilson, D. (1985/6), 'Loose talk'. Proceedings of the Aristotelian
Society LXXXVI: 153-71. Reprinted in Davis 1991: 540-9.
Sperber, D. and Wilson, D. (1987a), 'Precis of Relevance'. The Behavioral and
Brain Sciences 10.4: 697-710.
Sperber, D. and Wilson, D. (1987b), 'Presumptions of relevance'. The Behaviorial and Brain Sciences 10.4: 736-54.
Sperber, D. and Wilson, D. (1990a), 'Spontaneous deduction and mutual
knowledge'. The Behavioral and Brain Sciences 3.1: 179-84.
Sperber, D. and Wilson, D. (1990b), 'Rhetoric and relevance', in Bender and
Wellbery 1990: 140-56.
Sperber, D. and Wilson, D. (forthcoming a), 'Fodor's frame problem and
relevance theory (reply to Chiappe and Kukla)'. To appear in The Behavioral
and Brain Sciences.
Sperber, D. and Wilson, D. (forthcoming b), Relevance and meaning. Blackwell,
Oxford.
Stainton, R. (1993), Non-sentential assertions. MIT, Ph.D. dissertation.
Stainton, R. (1994), 'Using non-sentences: An application of relevance theory'.
Pragmatics and Cognition 2.2: 269-84.
Stalnaker, R. (1974), 'Pragmatic presuppositions', in Munitz and Unger 1974:
197-213.
Stalnaker, R. (1978), 'Assertion', in Cole 1978: 315-32.
Steinberg, D. and Jakobovits, L. (eds) (1971), Semantics: An interdisciplinary
reader. Cambridge University Press, Cambridge.
Stenning, K. (1978), 'Anaphora as an approach to pragmatics', in Halle, Bresnan
and Miller 1978: 162-200.
van Straaten, Z. (ed.) (1980), Philosophical subjects. Clarendon Press, Oxford.
Strawson, P. (1964a), 'Intention and convention in speech acts'. Philosophical
Review 73: 439-60. Reprinted in Searle 1971: 170-89.
Strawson, P. (1964b), 'Identifying reference and truth values'. Theoria 3:
96-118.
Strawson, P. (1969), 'Meaning and truth'. Inaugural lecture at the University of
Oxford. Reprinted in Strawson 1971: 170-89.
Strawson, P. (1971), Logico-linguisticpapers. Methuen, London.
Sun, Yu (1993), 'Pragmatic inference in relevance theory'. Waiguoyu 4.86:
39-43.
Swinney, D. (1979), 'Lexical access during comprehension: (Re)consideration
of context effects'. Journal of Verbal Learning and Verbal Behavior 18.6:
645-60.
Taglicht, J. (1984), Message and emphasis: On focus and scope in English.
Longman, London.
Tanaka, K. (1992), 'The pun in advertising: A pragmatic approach'. Lingua 87.
1/2: 91-102.
Tanaka, K. (1994), Advertising language: A pragmatic approach to advertisements in Britain and Japan. Routledge, London.
Tanenhaus, M. and Lewman, J. (1979), 'Evidence for multiple stages in the
Bibliography
317
processing of ambiguous words in syntactic contexts'. Journal of Verbal

Learning and Verbal Behavior 18: 427-40.
Taylor, J. (1993), 'Possessives and relevance'. Stellenbosch Papers in Linguistics
26: 14-34.
Taylor, T. and Cameron, D. (1987), Analyzing conversation. Pergamon Press,
New York.
Thomason, R. (1970), Symbolic logic. Macmillan, London.
Tirkkonen-Condit, S. (1992), 'A theoretical account of translation: Without
translation theory?' Target 4.2: 237-45.
Todorov, T. (1977), Theories du symbole. Seuil, Paris.
Toolan, M. (1992), 'On relevance theory', in Wolf 1992: 146-62.
Travis, C. (1981), The true and the false: The domain of the pragmatic. John
Benjamins, Amsterdam.
Travis, C. (1985), 'On what is strictly speaking true'. Canadian Journal of
Philosophy 15: 187-229.
Travis, C. (ed.) (1986), Meaning and interpretation. Blackwell, Oxford.
Travis, C. (1990), 'Critical notice of Relevance'. Canadian Journal of Philosophy
2.
Trotter, D. (1992), 'Analysing literary prose: The relevance of relevance theory'.
Lingua 87.112: 11-27.
Tsohatzidis, S. (1994), Foundations of speech act theory: Philosophical and
linguistic perspectives. Routledge, London.
Tyler, L. and Marslen-Wilson, W. (1977), 'The on-line effect of semantic
context on syntactic processing'. Journal of Verbal Learning and Verbal
Behavior 16.6: 683-92.
Uchida, S. (forthcoming), 'Text and relevance'. To appear in Carston, Uchida
and Song forthcoming.
Unger, C. (1994), 'The scope of discourse connectives and its relation to the
utterance in which it occurs'. University College London, M.A. dissertation,
to appear in Journal of Linguistics.
Vandepitte, S. (1989), 'A pragmatic function of intonation'. Lingua 79: 265-97.
Vandepitte, S. (1993), A pragmatic study of the expression and the interpretation
of causality: Conjuncts and conjunctions in modern spoken British English.
Koninklijke Academie voor Wetenschappen, Letteren en Schone Kunsten van
Belgie, Brussels.
Verschueren, J. and Bertuccelli-Papi, M. (eds) (1987), The pragmatic perspective.
John Benjamins, Amsterdam.
Verschueren, J., Ostman, J.-O. and Blommaert, J. (forthcoming), Handbook of
Pragmatics: Manual. John Benjamins, Amsterdam.
Vicente, B. (1992), 'Metaphor, meaning and comprehension'. Pragmatics 2:
49-62.
Vygotsky, L. (1962), Thought and language, translated from the Russian (1934).
MIT Press, Cambridge, MA.
Walker, R. (1975), 'Conversational implicatures', in Blackburn 1975: 133-81.
Walker, R. (1989), 'Review of Relevance'. Mind and Language 4.112: 151-9.
Watson, R. (1995), 'Relevance and definition'. Journal of Child Language 22:
211-22.
318
Bibliography
Watts, R. (1988), 'A relevance-theoretic approach to commentary pragmatic

markers: The case of actually, really and basically'. Acta Linguistica Hungarica 38.1-4: 235-60.
Werth, P. (ed.) (1981), Conversation and discourse. Croom Helm, London.
Wilks, Y. and Cunningham, C. (1986), 'A purported theory of relevance', in
Mey 1986: 383-418.
Wilson, D. (1975), Presuppositions and non-truth-conditional semantics. Academic Press, London. Reprinted in Gregg Modern Revivals in Philosophy.
Gregg Revivals, Aldershot (1991).
Wilson, D. (1992a), 'Reply to Chametzky'.Journal of Pragmatics 17: 73-7.
Wilson, D. (1992b), 'Reference and relevance'. VCL Winking Papers in Linguistics 4: 165-91.
Wilson, D. (1994a), 'Relevance and understanding', in Brown et al. 1994: 35-58.
First published in Guijarro Morales 1993.
Wilson, D. (1994b), 'Truth, coherence and relevance'. Paper delivered to the
European Society for Philosophy and Psychology, July.
Wilson, D. and Smith, N. (eds) (1992), 'Special issue on relevance theory
(volume 1)'. Lingua 87.112.
Wilson, D. and Smith, N. (eds) (1993), 'Special Issue on relevance theory
(volume 2)'. Lingua 90.112.
Wilson, D. and Sperber, D. (1981), 'On Grice's theory of conversation', in
Werth 1981: 155-78.
Wilson, D. and Sperber, D. (1986a), 'Inference and implicature', in Travis 1986:
43-75. Reprinted in Davis 1991: 377-93.
Wilson, D. and Sperber, D. (1986b), 'Pragmatics and modularity', in Parasession
on pragmatics and grammatical theory. Chicago Linguistics Society 22: 67-84.
Reprinted in Davis 1991: 583-95.
Wilson, D. and Sperber, D. (1986c), 'An outline of relevance theory', in Alves
1986: 19-42. Reprinted in Notes on Linguistics (1987) 39: 5-24.
Wilson, D. and Sperber, D. (1987), 'The self-appointment of Seuren as censor:
A reply to Pieter Seuren' . Journal of Semantics 5: 145-62.
Wilson, D. and Sperber, D. (1988a), 'Mood and the analysis of non-declarative
sentences', in Dancy, Moravcsik and Taylor 1988: 77-10l.
Wilson, D. and Sperber, D. (1988b), 'Representation and relevance', in Kempson
1988a: 133-53.
Wilson, D. and Sperber, D. (1992), 'On verbal irony'. Lingua 87.112: 53-76.
Wilson, D. and Sperber, D. (1993), 'Linguistic form and relevance'. Lingua
90.112: 1-25.
Wilson, D. and Sperber, D. (forthcoming), 'Pragmatics and time'. To appear in
Carston, U chida and Song forthcoming.
Wilson, J. (1990), Politically speaking: The pragmatic analysis of political
language. Blackwell, Oxford.
Winckler, W. and van der Merwe, C. (1993), 'Training tomorrow's bible
translators: Some theoretical pointers'. Journal of Northwest Semitic
Languages 19: 41-58.
Winograd, T. (1977), 'A framework for understanding discourse', in Just and
Carpenter 1977: 72-86.
Bibliography
319
Wolf, G. (ed.) (1992), New departures in linguistics. Garland, New York.
W right, R. (1975), 'MeaningNN and conversational implicature', in Cole and
Morgan 1975: 363-82.
Yoshimura, A. (1993a), 'Explicature and implicature formation in the modeling
of metaphor and metonymy'. Osaka University Papers in English Linguistics
I: 175-84.
Yoshimura, A. (1993b), 'Pragmatic and cognitive aspects of negative polarity'.
Osaka University Papers in English Linguistics I: l41-73.
Yu, P. (1979), 'On the Gricean program about meaning'. Linguistics and
Philosophy 3.2: 273-88.
Zegarac, V. (1991), Tense, aspect and relevance. University of London, Ph.D.
thesis.
Zegarac, V. (1993), 'Some observations on the pragmatics of the progressive'.
Lingua 90.1/2: 201-20.
Ziff, P. (1967), 'On H. P. Grice's account of meaning'. Analysis 28: 1-8.
Reprinted in Steinberg and Jakobovits 1971: 60-5.
Ziv, y. (1988), 'On the rationality of "Relevance" and the relevance of
"Rationality"'.Journal of Pragmatics 12.5/6: 535-45.
Index
Abelson, R., 285 n14

accessibility, 77, 78, 81, 138-51, 160-1, 167-71,234
advertising, 297 n28
advice, 231,250-1
Allerton, D., 288 n12
ambiguity, 10, 13, 34,168-9,175,188
argumentation, 298 n28
Ariel, M., 296 n15, 297 n21
Armstrong, D., 281 nIl, 282 nl7
Arnauld, A., 5-6
artificial intelligence, 93-103, 129-30, 131,265,298
n29
asking, 246-7, 252-3
assertions, 180-1,202,214-15,245
ordinary, 181, 183, 193,202,231,242,248,249
assumption, 2, 39-46, 58-60, 68, 85-6, 138-9,
151-5,182,199-200,229-30
factual, 74-5, 81-3, 103-4,285 n7
schema, 73, 81-2, 167, 181-2, 188, 189,225,226,
285 nl0
strength of, 75-83, 108-17, 127-9,220-1,290
n30
Atlas, J., 296 nnl6, 17
attitude, 239-40
propositional, 10-11, 12, 57, 73-5, 180-1, 183,
230,246-54
Austin, J., 243
Austin, J. P. M., 296 n9, 297 n28
Bath,K., 17,20,69-70,251-2,281 nn7, 11,283 n27,
287 n2, 288 n3, 289 n23, 290 n27, 293 nla, 296
nnl4, 16, 17
Barthes, R., 8, 57
Bennett, J., 281 nIl, 282 n17
Berg, J., 295 n8
Bertolet, R., 296 n17
Bertuccelli-Papi, M., 296 n17
Bierwisch, M., 281 n6
Bird, G., 297 n22
Blackburn, S., 281 n 11
Blakemore, D., 258, 288 n4, 289 n21, 295 nn4, 6, 7,
9,296 n17, 297 nnl8, 21, 23, 26
Blass, R., 289 nnl9, 21, 295 n9, 296 n15, 297 nn21;
23,26
Booth, W., 294 n4c
Brockway, D., 289 n21
Brown, G., 70,216-17,287 n4, 288 n12
Burton-Roberts, N., 295 n8, 297 n26
Cairns, H., 288 n6
Cameron, D., 295 n8
Campbell, J., 296 n9, 298 n28
Cara, F., 260
Carlson, T., 283 n22
Carnap, R., 79-80
Carston, R., 257-8, 286 nnl6, 19,288 n4, 295 n7,
296 nnlO, 16,297 nnl8, 26, 298 nn32, 33, 34
categories, logical and syntactic, 205-6
central thought processes, 47-8, 65-7, 71-5, 83-5,
90,151-5,165-6,185-7,293 nn2a, 2b
Chafe, W., 288 n12
Chametzky, R., 295 n8
Charolles, M., 295 n8
Chiappe, D., 295 n3
Chomsky, N., 8,94,196
Clark, B., 296 n9, 297 nn22, 26
Clark, H., 17-20, 283 n22, 284 n29, 288 nn9, 12,
290 n25, 296 nU
code, 3-15,24-8, 36-8, 43, 52-3, 63, 65, 166, 167,
169-70,172-6,177,191-3,282 n16
code model of communication, 2-21, 32, 44-6, 58,
166,169-70,174,196,230-1
cognitive effect, 263-6
cognitive environment, 38-46, 58-60, 62-4,151,
197,198,200,201,217-18,224,284 n32
coherence, 289 n 19
cohesion, 289 n19
Coleridge, S., 57
common knowledge, see mutual knowledge
communication
explicit, 11,37,55-60,162-3,174-6,182-93,
217-24,248,249,256-8,298 n34
implicit, 11, 12,34-7,55-60,162-3,174-5,
322
Index
182-3,193-202,217-24,235-7,239-42,
272-8, 298 nn30, 34
overt, 30-1, 60-4, 282 n20
strong, 59-60, 174-6
weak, 59-60, 212
see also code model of communication;
ostensive-inferential communication
concept, 85-93, 204
classificatory, 79, 124, 144, 152, 182
comparative, 79-81, 124-9, 144-5, 152-3, 182
quantitative, 79-81, 129-32
theoretical, 119-20, 125
conceptual-procedural distinction, 258, 297 n21;
see also constraints on relevance; semantics,
procedural confirmation, 65-6, 68-9, 75-82,
109-17, 164-72,201,211
see also strengthening
connectives, logical, 37, 84-103, 183,288 n4
constraints on relevance, 258-9, 289 n21, 297 nn22,
23
context, 12, 15-21, 36-7, 67, 89, 103, 118-23, 125,
129,132-51,163 n21, 167, 185-7,209,216-17,
221-2,223,236-7,287 n7
contextual effect, 108-37, 142-51, 156-7,202,204,
209-10,233-7,263-6,269-70,286 n26, 294
n2d
contextualisation, 107, 108, 113-16,286 n16
contradiction, 95, 114-15, 120-1, 148, 294 n2d
convention 25-7,52-3,217,226,230-1,249,282
n16
co-operative principle, 33, 35, 36, 161-2, 256, 260,
267-8, 273-4, 294 n3d, 296 n 14
Cormack, A., 288 n4
Cruttenden, A., 288 n12
Culpeper, J., 295 n8
Cunningham, c., 295 n8
Davidson, D., 281 nil
Davies, M., 281 nil
Davis, S., 296 nl0
de B"eaugrande, R., 70
declarative, 246-7
see also mood
deductive device, 93-108, 110-17,126,137-40,
145, 286 n20, 287 n7, 293 n2b
description, 224-31, 232, 247-8, 289 n22
descriptive use, 224-31, 232, 247-8, 259, 289 n22
Dinneen, D., 288 n12
disambiguation, 12, 13,34, 168-9, 175, 179,183-7,
190-1,204-8,256-7,288 n6, 296 n14
discourse markers, 258-9, 297 n22
Dressier, W., 70
Dretske, F., 2
Ducrot, 0., 281 n7, 297 nn20, 21, 26
Durant, A., 297 n27
echoic interpretation, 238-9, 259, 297 n26
effability, principle of, 191-2
efficiency, 46-9, 152, 167-8,204,263-6,289 n21
elimination rule, 86-93, 97-103
and-elimination, 86
conjunctive modus ponens, 99, 100
disjunctive modus ponens, 99, 100
modus ponendo ponens, 87,98-9,100,105
modus tollendo ponens, 87

encyclopaedia, 87-90, 92-3,107,185-6,190-1,
197-8,204,236-7,286 n17
enrichment, 181, 183, 184, 188-91, 204, 256-8, 279,
296 n17, 298 n34
entailment, 84, 95-6, 101-2
epizeuxis, see repetition
Escandell Vidal, V., 295 n8
Espinal, T., 296 nn9, 17,297 n23
evidence, 22-3, 25-6, 51-3, 64, 75-83,108-17,
164-5,249,250
evolution, 260, 261-3
exclamative, 231, 253-4, 259
experimental tudies, 183-8, 190-1,257,259,283
n22, 285 n9, 286 n21, 288 n6,
explicature, 182-3, 193, 224, 244, 290 n30, 294 n4a
constraints on, 258-9
explicit-implicit distinction, 176-83,256-8,298
n34
Fabb, N., 297 n27
Ferrar, M., 296 n9, 297 n28
figurative language, 11, 12, 56-8, 166,219-43,283
n27
focal scale, 208-12, 214-15
focally stressed constituent, 203, 210
focus, 202-17, 288 n17
Fodor, J. A., 66-7,71,91-2,117,177,186,257,
263,283 n28, 285 nn3, 12, 13,286 nn15, 18,
287 ni, 288 n6, 293, n2b
Fodor, J. D., 286 n18, 289 n18
Forceville, c., 295 n9, 297 nn24, 28
Forget, D., 296 n15
Foster -Cohen, S., 260, 296 n 15
Fowler, A., 295 n3
frame, 88, 138
see also encyclopaedia
Fretheim, T., 296 n15, 297 n26
Frisch, K. von, 5
Frith, U., 260
Gabbay, D., 297 n21, 298 n29
gapping, 222-4
garden-path utterance, 184,242
Gardiner, A., 282 n12
Garnham, A., 293 nla, 295 n3
Garrett, M., 285 n13, 286 n18
Gazdar, G., 36-7,196,257,287 n3, 289 n18
Gerrig, R., 290 n25, 293 nla
Gibbs, R., 293 nla, 295 n8, 297 n24
Giora, R., 295 n8
Girotto, V., 260
Givon, Ri., 288 n12
Glucksberg, S., 297 n24
goals,
of cognition, 46-9, 263-6
Index
of communication, 161-2,267-8
Goatly, A., 297 n24
Good, D., 287 n3
Goodman, N., 69
Gorayska, B., 295 n8
Gordon, D., 36
grammar, generative, 8, 9-15, 94, 172-6,212-13
Green, G., 203, 288 n9, 294 n3d
Green, K., 297 n27
Grice, H.P., 2, 14,21-38,53-4,61,64, 161-3,
182-3,195-6,200-1,240-1,256,267-8,273,
281 nn7, 11,282 nn14, 17, 19,283 nn23, 24,
284 n30, 287 n10, 296 n12, 297 n19
see also maxims of communication; speaker's
meaning
Grimshaw, J., 253
Groefsema, M., 296 nn9, 17,297 n22
Grundy, P., 295 n8
guessing, 290 n30
Guijarro Morales,J.-L., 296 n10
Gundel, J., 296 n15
Gussenhoven, C, 288 n17
Gutt, E.-A, 259, 295 n7, 296 ng, 297 nn21, 22
Haegeman, L., 296 n17, 297 n21
hair-drier example, 30, 31, 61, 62-3
Halliday, M., 288 n12
Hamamoto, H., 297 n24
Happe, F., 260, 296 n9
Harman, G., 281 n11
Harnish, R.M., 17,20,69-70,251-2,281 nn7, 11,
283 n27, 287 n2, 288 nn3, 10,289 n23, 290
n27, 296 n14, 297 n22, 298 n32
Haviland, S., 283 n22, 288 n12, 296 n13
Hawkins, J., 296 n15
Higashimori, 1., 297 n21
Hinkelman, E., 293 n1a
Hirst, D., 295 n3, 296 n17
Hjelmslev, L., 7
Hobbs, J. 287 n6
Hogaboam, T., 288 n6
Horn, L., 296 n17, 298 nn30, 32
Hugly, P., 288 n10
humour, 297 n28
Hymes, D., 297 n24
hyperbole, 235-6
Ifantidou, E., 296 n9, 297 nn23, 26
illocutionary force, 10-11,226,243-54
imperative, 243, 246-7, 249-51, 254;
see also mood
implication
analytic, 104-8,204,215
background,209-11,214-17
contextual, 107-8, 112, 120, 126, 137, 143, 147,
148,204,221-2,233-5,275-6
foreground, 209-11, 214-17, 289 n20
logical,84
non-trivial,97-103
synthetic, 104-8, 112, 113, 137, 139
323
trivial, 97, 102-3, 106,284 n1

implicature, 34-7, 56-7, 69-70, 93, 133-7, 162-3
166,182-3,193-202,212,217-24 235-7 '
239-42, 244-5, 257-8, 275-6, 288 nn4, 1'0,
296 nn13, 14, 298 n34
bridging, 296 nn13, 14
calculability of, 200-1
constraints on, 258-9
see also semantics, procedural
determinacy of, 195-200, 235-7
generalised, 257
quantity, 36-7, 272-8, 298 nn 30, 34
scalar, 36-7, 272-8, 298 nn 30, 34
strength of, 197-202, 221-2
inference, 12-15, 16,36-8,57,68,75,137,173,
179-80,283 n28 deductive (demonstrative),
40,65,68-70,83-117,139-42,284 nI, 285
n11, 286 nn21, 24, 287 n2, 293 n2b
non-demonstrative, 37, 40, 65-71, 81-3,108-17,
201,254,284 n2
inferential model of communication, 2-3,12-15,
21-38,50-4,75, 104-7,282 n16
see also ostensive-inferential communication
information, 2, 88-90
given, 202-4, 210, 217
new, 48,107-17,118-19,132,137-51,202-4,
210,217,286 n25, 289 n20
old, 48,107-17,118-19,132,137-51
informativeness, 272-8, 298 n30
input (perceptual) processes, 65-6, 71-2, 81-2,
83,90,107,151-5,177-8,185-7,261-3,283
n28,293,nn2a,2b
intention, 21-38, 50-1, 58-9,153-5,165,169,
183,282 nn17, 18, 283 n21
communicative, 29-31, 60-4, 163-72, 179
informative, 29-31, 54-60, 61-2, 163-72, 179,
230-1,254,271-2,284 n32
interpretation, 224-43, 252-4, 289 nn22, 25
echoic, 238-9, 297 n26
interpretive use, 224-31, 237-43, 259, 297 nn 25, 26
interrogative, 231, 243, 246-7, 251-3, 254, 259
introduction rule, 95-103
and-introduction, 96, 98
double negation, 96
or-introduction, 96, 98-9, 100
irony, 11, 180,200-1,224-5,231,237-43,259,289
n25, 290 n26, 297 n24
ltani, R., 296 n9, 297 nn21, 23, 26
Jackendoff, R., 288 n12
Jakobson, R., 57, 281 n2
Jayez, J., 295 n3
Jiang, Y., 297 n21, 298 n29
Jodlowiec, M., 296 n9, 297 n28
Johnson-Laird, P., 102-3,281 n10, 285 n11, 286
nn18, 24, 287 nn2, 6, 288 n13
J orgensen, J., 290 n25
Jucker, A., 297 nn21, 28
Kahneman, D., 285 nn8, 9
Kamerman, J., 288 n6
324
Index
Kandolf, c., 296 n17

Kasher, A., 296 nl0, 297 n22
Katz, J., 191-3,283 n24, 285 n13, 286 n18, 290 n27
Kempson, R., 260, 283 n27, 288 n4, 289 nnl8, 21,
295 n7, 296 nnl0, 15,297 n21, 298 n29
Kerbrat-Orecchioni, c., 284 n31, 290 n25
Kiefer, F., 281 n6
Kiparsky, P., 260, 297 n27
Kleiber, G., 296 n15
Klinge, A., 297 n17
Ki:inig, E., 297 n23
Kreuz, R., 297 n24
Kripke, 5., 91,286 n15
Kukla,A., 295 n3
Lakoff, G., 36
Lancelot, c., 5-6
language, 8,9-15,16,52-3,172-7,191-3,196,284
n30, 285 n7, 298 n29
see also code
Leach, E., 7
Leech, G., 69, 281 n7, 183 n25, 295 n7
Lemmon, E., 93
Leslie, A., 295 n3
Levinson, 5., 69, 243, 276, 281 nn6, 7,283 n27, 287
n4, 295 n3, 296 nnl4, 16,298 nn30, 32, 33
Levi-Strauss, c., 8, 281 n3
Lewis, D., 18,26,281 nn9, 11,284 nn29, 30, 289
n23
Lewman, J., 288 n6
lies, 264
Lindsey, R., 295 n8
literal language, 11,25,191-2,230-7,239,251
literary studies, 8, 57, 75, 259-60, 265, 278-9, 297
n27
litotes, 238
Loar, B., 281 nll
Loftus, E., 281 n8
logical form, 72-5, 81-2, 85-6, 179-80, 189, 191-3,
252-3,285 nn4, 7
development of, 181-2
loose talk, 233-7, 279
Lucy, P., 283 n22
Lunn, P., 297 n22
Luscher, J.-M., 297 n21
Lyons, J., 281 n2, 287 n4, 288 n12
McCawley, J., 93, 287 n6, 293 nla
McDowell, J., 281 nIl
oMacLaran, R., 289 n21
manifestness, 39-46, 58-60, 151-5, 163-5,
196-200, 283 n28
mutual, 41-6, 60-4, 163-5, 284 nn29, 32, 293
nla
Mao, R., 296 n9
Marshall, c., 17-20,283 n22, 284 n29, 296 n13
Marslen-Wilson, W., 287 nI, 288 n6
Martin, R., 297 n24
Matsui, T., 296 n9, 296 nn13, 15
maxims of communication, 13-14,33-8,161-3,
179-80,200-1,230-1,249,271,273-8,283
nn24, 26, 287 n2, 296 n14, 298 n30
Mayher, J., 298 n28
meiosis, 238
mental model, 285 nIl, 286 n24, 287 n2
mention, 289 n25
metaphor, 180,224-5,231-7,242-3,248,251,259,
279, 283 n27, 290 n26, 297 n24
metonymy, 237
Mey, J., 295 n3
Miller, G., 286 n18, 164 n25
Minsky, Mo, 285 n14
Mitsonobu, M., 296 nIl
modularity, 65-7,71-2,83,177-8,260,293 nn2a, 2b
Moeschler,J., 295 nn6, 7,296 nn9, 10,297 nnl7,
21, 22, 23, 26, 298 n28
mood, 73, 180,247,297 nn22, 26
imperative, 73, 247, 254, 297 n25
indicative, 73, 180,247,254,297 n26
subjunctive, 247
Morgan, J., 288 n9, 294 n3d
Morris, c., 281 n6
mutual knowledge, 15-21, 38-46, 179-80,281 nl0,
283 n21, 284 nn29, 2, 293 nla
see also mutual manifestness
Nasta, D., 296 n9, 298 n28
Neale, 5., 296 nnl2, 16
Nebeska, 1., 295 n8
Neisser, U., 281 n8
Nishiyama, y., 295 n8
N"lke, H., 297 n23
Oh, C.-Ko, 288 n12
O'Neill, J., 295 n8
optimal processing, 144, 152-3, 154, 156-7
ostension, 49-54, 71, 155
ostensive-inferential communication, 23, 49, 50-4,
60-4,65-6,153-72,174,176,226-31,244-5,
271-2,275
see also inferential model of communication
parallelism, 213, 222-4
Parkes, c., 285 n13, 286 n18
parsing, 204-8
Pateman, T., 295 n3
Patton, T., 281 nIl
Peirce, c., 6
perceptual processes, see input processes
Perfetti, c., 288 n6
Pemer, J., 293 nla, 295 n3
Perrin, L., 296 n9, 297 n24
phenomena, 40,151-5,160-1
Pike, K., 7
Pilkington, A., 296 n9, 297 nn24, 27
poetic effects, 196,217-24,235-7
politeness, 297 n28
Politzer, G., 260, 295 n3, 296 n9
Posnanski, V., 296 n9, 298 n29
pragmatics, 10-15,20-1,36-8,56-60,69-71,
Index
141-2,179-80,182-3,194-6,201-2,217,
230-1,243-7,281 nn6, 7, 289 n21
Gricean, 21-38, 60-4, 161-3, 182-3, 195-6,
200-1,240-1,256,267-8,273-8,296 n13
predictions, 245
presumption of relevance, 156-9, 161, 164-72,234,
266-78,287 nIl, 294 n3c, 298 n31
presupposition, 202, 203, 213-15, 217, 283 n27, 287
n6, 289 nnl8, 21
Prince, E., 288 n12
principle of relevance, 50,158-63,166-72,178,
184-5,195,196-7,201,220-2,229,231,
233-5,248,249,251,253,254,262-78,287
nIl, 293 nlc, 294 n3d, 294 n4b
processing effort, 46-50,124-32,136-7,142,
149-50,156-7,166-8,185,190-1,196-7,204,
213,216,220-2,223,233-6,268-9
promises, 245, 246, 290 n28
propositional form, 72-5,179-93,228,231-7,
252-3
see also assumption; logical form
prototype, 88,91-2, 138
Pulman, S., 286 n16
Pumam, H., 91, 286 n15
questions, 194,251-3
indirect, 251-2
relevant, 203, 207-8, 214, 215
rhetorical, 247, 251-2
Wh-, 203, 207-8, 252-3
yes-no, 225, 252-3
Quine, W., 288 n8
quotation, 227-8
see also echoic interpretation; mention
Reboul, A., 295 nn6, 7, 296 nn9, 15,297 nn22, 24,
27
Recanati, .,281 n11, 282 n19, 290 n27, 296 nnl5,
16,297 nnl7, 22, 24
Reddy, R., 281 nl
reference assignment, 10, 12, 17-21, 91, 175, 179,
183-4,187-8,190-4,204-8,214-15,256-7,
288 n7, 296 nn13, 14, 15
Reinhart, T., 288 n12
relevance, 36, 38, 46-50, 51-2, 70-1, 75, 103, 117,
. 118-72,209-10,238,252-3,260-78,298 n30
m a context, 123-132, 142, 265
and evolution, 261-3
to an individual, 142-51,265-6,294 n3a
maximal, 261-3, 293 nlc, 298 n30
optimal, 157-60, 194, 196,204,208,214,218,
222,234,266-72,287 nIl
of phenomena, 151-5,261-3,287 n8, 294 n3a
of stimuli, 153-5
to a topic, 216-17
and truth, 263-6, 293, nIb
see also principle of relevance
reminders, 137, 143, 149-50,211
repetition, 219-22
325
reported speech, 224-31, 248-9, 250

reported thought, 224-31
representation, 89-90, 224-31
conceptual, 2, 27, 65,72-5,81-3,85 131-2
151-5,231,258
"
descriptive, 227-9
interpretive, 229-31,252-4
semantic, 9,11,36,174-6,177,179 185-6 188
191-3,257-8
"
,
sensory, 2, 71-2, 81, 151-5, 283 n28
requests, 60-4,225,231,250-3
resemblance, 226-37, 289 n25
Richards, c., 297 n27
Richards, I., 57
Rips, L., 100, 102, 286 nn21, 22
Roberts, L., 295 n8
Rochemont, M., 288 nnl0, 14, 17
Rouchota, V., 296 nn9, 15,297 n22
Russell, S., 293 nla
Sadock, J., 283 n27, 288 nl0, 290 n29, 295 n8
Sag, I., 287 n6
Sanchez de Zavala, V., 295 n8
Saussure, F. de, 6, 7, 8
saying, 180-1,246-9,253-4
Sayward, c., 288 n 10
Scancarelli, J., 297 n17
scenario, 138, 182
Schank, R., 285 n14
schema, 88, 138, 175, 190-1,236
see also assumption; logical form
Schiffer, S., 17, 18,21,24,30,31,60-1,281 nn9,
11,282 nn 14, 18, 283 n21
Schlegel, A. and F., 57
Schunk, D., 283 n22
script, 88
Searle, J., 24, 25-6, 243-4, 251-2, 281 nn6, 11, 282
nnl4, 17,20,283 n27, 289 n23
semantics, 172-3, 193,224-31
of conceptual representations, 257-8
formal, 257
linguistic, 193, 257-8, 258-9
non-truth-conditional, 257, 258-9, 297 nn20, 21,
22
procedural, 258-9, 289 n21, 297 nn21, 22, 23
speech-act, 258-9, 297 nn22, 23
truth-conditional, 224-31, 257
semantics-pragmatics distinction 9-15 176-93
257-8
'
,
,
semiology, see semiotics
semiotics, 6-21, 24, 55, 57, 196,202,281 n5
Seto, K.-I., 297 n24
Seuren, P., 295 n3
Shannon,C.,4,5
Sinclair, M., 295 nn6, 8
Slovic, P., 285 n9
Smith, A., 297 n21
326
Index
Smith, N., 260, 281 n4, 289 n21, 295 n7, 296 nl0,
297 nn21, 23
Soames, S., 289 n18
social factors in communication, 30-1, 60-64, 279,
297 n28
Song, N.-S., 297 n24
speaker's meaning, 21-38, 53-4, 56-7, 281 nll
speech acts, 10-11,224-6,243-54,258,259,290
nn27, 28, 297 nn 22, 25
assertive, 243-4, 246-7
directive, 244, 246-7, 251-2
indirect, 244, 283 n27
see also assertions; promises; questions;
reminders; saying; telling; thanking; threats;
warnings
Sperber, D., 258, 260, 281 nn3, 4, 10, 283 nn24, 26,
285 nn4, 6, 288 nl0, 289 nn22, 25, 293 nnla,
2a, 294 nn3d, 3e, 295 nn4, 5, 7, 296 nn13, 16,
297 nn21, 22, 23, 24, 27
Stainton, R., 296 n9, 297 n17
Stalnaker, R., 283 n27, 287 n6, 289 n21
Stampe, D., 281 nIl
Stenning, K., 287 n6
stimuli, 2, 29,58,61,63,67,81,132,160,163,164,
177
ostensive, 153-61, 166, 168-9, 174...:.5, 189,
226-7,254
relevance of, 153-5
Strawson, P., 21, 23, 24, 28-30, 60-1, 214-15, 281
nll, 283 n21
strengthening, 120, 147-8
dependent, 112, 114
independent, 112-14, 143
retroactive, 115-17, 190-1
see also assumption; confirmation
stress, 202-3, 209-17, 288 nB
contrastive, 212-3, 288 nnll, 16
stylistic effects of, 196, 202-24
Sun, y., 295 n8
Swinney, D., 288 n6
symbolism, 281 n5
synecdoche, 237
Taglicht, J., 288 n12
Talbot, M., 295 n3
Tanaka, K., 296 n9, 297 n28
Tanenhaus, M., 288 n6
Taylor, J., 297 n17
Taylor, T., 295 n8
telling, 246-7, 249-51, 294 n4a
thanking, 245, 246, 290 n28
theme, 202, 215-16
Thomas, J., 295 n7
Thomason, R" 93
thoughts, 29,9,10,149,191-3,200,229-31,233-5,
238-9, 252-4, 288 n8
see also central thought processes
threats, 245
Tirkkonen-Condit, S., 297 0.25
Todorov, T., 6
Toolan, M., 295 n8
topic, 202, 216-17
translation, 228, 259, 297 n25
Travis, c., 288 n4, 295 n3
tropes, 224-43, 259, 297 n24
Trotter, D., 297 n27
truth conditions, 214, 227, 296 n14
see also description
truth and relevance, 263-6
Tsimpli, 1., 260
Tversky, A., 285 nn8, 9
Tyler, L., 288 n6
Uchida, S., 297 n27
Unger, c., 297 n21
van der Merwe, c., 297 n25
Vandepitte,S., 296 n9, 297 n21
Vicente, B., 297 n24
Voltaire, 241, 294 n4c
Vygotsky, 1., 6
Walker, E., 285 n13, 286 n18, 295 n3
Walker, R., 283 n24, 288 nl0
warnings, 245
Watson, R., 260
Watts, R., 297 n23
Weaver, W., 4, 5
Wilks, Y., 295 n8
Wilson, D., 258, 281 nn4, 10,283 nn24, 26, 27, 288
nl0, 289 nn18, 25, 293 nla, 294 n3d, 295 nn4,
5,7,296 nnto, 13, 15, 16,297 nn20, 21, 22, 23,
24,27
Wilson, J., 298 n28
Winckler, W., 295 n6, 297 n25
Winograd, T., 285 n14, 288 n5
word meaning, 90-3
Wright, R., 281 nll
Yoshimura, A., 297 nn23, 24, 26
Yu, P., 281 nIl
Yule, G., 70, 216-17, 287 n4, 288 n12
Zegarac, V., 296 n9, 297 n23
zeugma, see gapping
Ziff, P., 281 nll
Ziv, Y., 295 n8
Zwicky, A., 290 n29

Sperber Wilson Relevance 1996 PDF

Uploaded by

Copyright:

Available Formats

Sperber Wilson Relevance 1996 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sperber Wilson Relevance 1996 PDF

Uploaded by

Copyright:

Available Formats

Relevance

Communication and Cognition

Copyright Dan Sperber and Deirdre Wilson, 1986, 1995

British Library Cataloguing in Publication Data

Library o/Congress Cataloging in Publication Data

Printed and bound in Great Britain by T. J. Press Ltd, Padstow, Cornwall

Preface to Second Edition

Non -demonstrative inference

4 Aspects of verbal communication

The identification of propositional form &

Notes to First Edition

Notes to Second Edition

Preface to Second Edition

Preface to Second Edition

programme we had merely intended to outline. The months became

How do human beings communicate with one another? For verbal

Certainly, what is communicated by a religious ritual is quite different

The code model and the semiotic approach

machines) to communicate. A message is a representation internal to the

The code model and the semiotic approach

The view of linguistic communication as achieved by encoding

The code model and the semiotic approach

Whenever communication is observed, an underlying system of signs is

By studying rites, customs, etc. as signs, I believe that we shall throw

Decoding and inference

2 Decoding and inference in verbal comprehension

As already mentioned, a generative grammar is a code which pairs

sentences belongs to grammar; the study of the interpretation of

Decoding and inference

However, most aspects of utterance interpretation cannot be handled so

Decoding and inference

Sentence (16) is ambiguous, and can be understood as conveying either

standards, and hearers systematically expect them to, a whole range of

(21) Premises: If P then Q

The mutual-knowledge hypothesis

Do speakers and hearers in ordinary verbal communication generally

The set of premises used in interpreting an utterance (apart from the

The mutual-knowledge hypothesis

are sometimes improperly applied, or because these mechanisms at best

the film is Monkey Business, and be referring to that. Or perhaps he might

The mutual-knowledge hypothesis

to avoid misunderstanding whenever possible' (Clark and Marshall1981:

(26) The door's open.

Speaker and hearer might have shared knowledge of hundreds of different

Grice, 'meaning' and communication

hypothesis can be empirically justified, we want to approach things the

4 Grice's approach to 'meaning' and communication

'CS] meant something by x' is (roughly) equivalent to '[5] intended

Grice, 'meaning' and communication

Mary's intentions in this example are quite similar in structure to those we

We have now looked at two models of communication. According to the

The code and inferential models

It may be true that most human communication involves the use of

The code and inferential models

generalis~d, all ~ecoding, lir:g~isti~ decoding included, is s~e~ as an

the recogmtion of the informative intention will itself lead to the

7 Problems of explanation: Grice's theory of conversation

terms 'meaning' or 'communication'. From our current, more psycholo_

bout what the communicator might have intended to convey: it is that