Language Mind and Brain Dabrowska Libros

DABROWSKA pre Make up 26/3/04 4:32 pm Page i
Language, Mind and Brain

DABROWSKA pre Make up 26/3/04 4:32 pm Page ii
To the memory of Mieczysl-aw Dabrowski

DABROWSKA pre Make up 26/3/04 4:32 pm Page iii
Language, Mind and

Brain
Some Psychological and Neurological
Constraints on Theories of Grammar
E WA DA
B ROWSK A
EDINBURGH UNIVERSITY PRESS

DABROWSKA pre Make up 26/3/04 4:32 pm Page iv
Ewa Dabrowska, 2004
Edinburgh University Press Ltd

22 George Square, Edinburgh
Typeset in 10/12 Times New Roman

by Servis Filmsetting Ltd, Manchester, and
printed and bound in Great Britain by
??????????????????????????????????????
A CIP record for this book is available from the British Library
ISBN 0 7486 1474 5 (hardback)

ISBN 0 7486 1472 3 (paperback)
The right of Ewa Dabrowska

to be identified as author of this work
has been asserted in accordance with
the Copyright, Designs and Patents Act 1988.
DABROWSKA pre Make up 26/3/04 4:32 pm Page v
Contents
List of figures viii

List of tables ix
Acknowledgements x
Chapter 1 Introduction 1
PART I THE BASIC SPECIFICATIONS 5
Chapter 2 Language processing: speed and flexibility 7

1. The complexity of the task 7
2. Speed 13
3. Processing shortcuts 17
3.1 Prefabricated units 18
3.2 Shallow processing 22
3.3 Frequency 25
4. Implications for linguistic theory 26
Chapter 3 Language acquisition: robustness 29

1. Individual differences 30
2. The role of input 33
3. Language development in blind children 36
4. The robustness of language 38
Chapter 4 Language in the brain 40

1. The localisation issue 41
2. Preservation of grammatical knowledge in Brocas aphasia 43
3. The co-occurrence of lexical and grammatical deficits 47
4. The resilience of language 48
Chapter 5 Language and other cognitive processes 50

1. The modularity hypothesis 50
2. Problems with the double dissociation argument 52
DABROWSKA pre Make up 26/3/04 4:32 pm Page vi
vi Language, mind and brain
Chapter 6 Biological underpinnings 58

1. A genetically specified language module? 58
2. Human adaptations to language 61
2.1 Preadaptations 61
2.2 Cultural learning and mind-reading 62
2.3 Cortical control over vocalisations and coordinating action
sequences 65
3. Language adaptations to humans 67
4. Universal Grammar again 69
4.1 Problems with the poverty-of-the stimulus argument 69
4.2 Some new questions 72
5. Conclusion 75
PART II THE BUILDING BLOCKS OF LANGUAGE 79
Chapter 7 Words 81
1. The semantics of locative terms 83
1.1 Conventional construal 85
1.2 Polysemy 89
1.3 Cross-linguistic variation 91
2. The acquisition of locative terms 100
2.1 Warlpiri 101
2.2 Tzotzil 101
2.3 Japanese 102
2.4 Korean 103
2.5 General discussion 104
3. Innate structure, yes but of what kind? 105
3.1 Semantic primitives 106
3.2 Perceptual primitives 107
4. Lexical learning in a constrained connectionist network 111
5. Conclusion 112
Chapter 8 On rules and regularity 116

1. Words and rules 116
2. The connectionist bombshell 117
3. The linguists strike back 118
4. The dual-mechanism model 119
5. The German plural: a minority default? 125
6. The Polish genitive: an inflectional system without a default 130
6.1 The system 130
6.2 Genitive masculine forms in default contexts 131
6.3 Adult productivity with genitive endings 135
6.4 Acquisition of the genitive inflection 138
7. The final test case: the Polish dative 144
DABROWSKA pre Make up 26/3/04 4:32 pm Page vii
Contents vii
8. Interim conclusions 147

9. Challenges for connectionism 149
Chapter 9 Syntactic Constructions 159

1. Ties between lexical and grammatical knowledge 159
2. Multi-word units in acquisition 161
2.1 Premature usage 161
2.2 Developmental U-curves 161
2.3 Inappropriate and ungrammatical usage 162
2.4 Pronoun reversals 163
2.5 Filler syllables 164
2.6 Lexically based patterns 166
2.7 Mosaic acquisition 173
3. A case study: questions 178
3.1 Early interrogative utterances 178
3.2 From formula to schema 180
3.3 Where do non-formulaic utterances come from? 185
3.4 Evidence for piecemeal learning 186
3.5 Questions with long-distance dependencies 196
3.6 Conclusion 200
Chapter 10 The cognitive enterprise 203

1. Towards a psychologically realistic grammar 203
2. A crash course in Cognitive Grammar 204
2.1 Linguistic expressions are symbolic units 204
2.2 Imagery 205
2.3 Things, processes and atemporal relations 207
2.4 Constructional schemas 208
2.5 Language as a structured inventory of conventional
linguistic units 210
2.6 A usage-based model 213
2.7 Meeting the specifications 214
3. Language production in a CG framework 214
4. A cognitive view of language acquisition 219
5. More on regularity 225
6. Future directions 227
Bibliography 230
Index 259
DABROWSKA pre Make up 26/3/04 4:32 pm Page viii
List of figures
7.1 Is the banana in the bowl? 85

7.2 mah away up there to the side in the face of the slope 88
7.3 Another virtual container 89
7.4 Some in, on and above relationships 90
7.5 Some ways of encoding on relationships in Tzeltal 95
7.6 Some ways of encoding in relationships in Tzeltal 96
7.7 Figure/ground ambiguities 97
7.8 A spatial scene and four observers 98
7.9 The Tzeltal geocentric system of projective relationships 99
8.1 Childrens performance on the nonce-word production task 139
8.2 Correct uses and overgeneralisations of genitive endings in the
four children 140
8.3 Cumulative number of masculine and feminine noun types used
in the genitive case 141
9.1 Formulaic and non-formulaic questions in the three children 181
9.2 Proportion of inverted Y/N questions 187
9.3 Contraction of auxiliary/copula is after what, where and who 193
10.1 The symbolic units []/[kt] (a) and [[]-]/[[kt]<[s]]
(b) 204
10.2 A simple clause (a), a low-level schema (b), and an abstract
schema (c) 209
10.3 Part of the semantic network for fly 212
10.4 Superimposition of a typical frame and filler 216
10.5 Superimposition with mutual elaboration 216
10.6 The emergence of a schema 221
DABROWSKA pre Make up 26/3/04 4:32 pm Page ix
List of tables
8.1 Circumstances in which memory is not accessed and the default

inflection is applied 121
8.2 Genitive inflections on masculine nouns in default
circumstances 136
8.3 Response types in the nonce-word production study 137
8.4 Genitive endings on masculine nouns in the Object, Place and
Substance conditions 137
8.5 Overgeneralisation errors in the developmental corpora 142
8.6 Case marking on noun lexemes used in genitive contexts 143
8.7 Response types in the dative study 146
9.1 Formulaic questions in the speech of five 2-year-olds 179
9.2 Data for the longitudinal study 180
9.3 The childrens 10 earliest formulas (in order of emergence) 182
9.4 Errors of commission in formulaic and non-formulaic questions 186
9.5 Inverted Y/N questions according to auxiliary 188
9.6 Questions with present-tense forms of the auxiliary do 189
9.7 Inverted Y/N questions according to person and number in
questions with auxiliaries other than do 190
9.8 Inverted and uninverted questions with first person singular
subjects 192
9.9 Child questions with long-distance dependencies in five other
CHILDES corpora 197
9.10 Examples of sentences used in the grammaticality judgement
study 199
9.11 Acceptability ratings 199
DABROWSKA pre Make up 26/3/04 4:32 pm Page x
Acknowledgements
I would like to thank my teachers Roman Kalisz and Tomasz Krzeszowski,

who first introduced me to cognitive approaches to language, and thus started
me on the long way that lead to this book. I am very grateful to Magdalena
Smoczynska for encouraging me to take up work on child language, and gen-
erously providing me with a coded version of the Krakw corpus to get me
going. The research on the acquisition of Polish case inflections reported in
Chapter 8 is based on this data.
I began working on the book while I was visiting the Max Planck Institute for
Evolutionary Anthropology in Leipzig. During the visit, and on many subse-
quent occasions, I had many interesting and fruitful discussions with Elena
Lieven, Mike Tomasello, Holger Diessel and Heike Behrens. Many of the ideas
in this book emerged and developed during these discussions, and I am extremely
grateful to all of them for their generosity and encouragement. Special thanks
are due to Mike Tomasello and Elena Lieven, who have read and commented on
parts of the manuscript and provided many valuable and insightful comments. I
am also indebted to Wojciech Kubinski, Ola Kubinska, Agata Kochanska, Anna
Tyminska and Beata Williamson for help with the data collection.
Most of the book was written during a period of leave funded by the
University of Sheffield and the Arts and Humanities Research Board (grant
number AN5597/APN15868). The research on case inflections reported in
Chapter 8 was supported by a British Academy Small Research Grant (grant
number RB 100556).
I would like to thank Mouton de Gruyter for permission to reproduce
figures 7.4, 7.5, and 7.6, which originally appeared as figures 1, 2, and 3,
respectively, in P. Brown, The INs and ONs of Tzeltal locative expressions:
the semantics of static descriptions of location, Linguistics, 32, 74390.
Most of all, I would like to thank my husband, Michael Pincombe, for
acting as linguistic consultant and voice of reason on topics related to the
book and many others. Michael read the entire manuscript and identified
many issues that needed clarification or correction; he produced the artwork
in Chapter 7; and his humour, good cheer and papriks have helped to keep
me happy and sane (well, almost) during the duration of the project.
DABROWSKA text Make up 26/3/04 4:34 pm Page 1
1 Introduction
Linguistics is often dened as the branch of cognitive science dealing with

language but, most linguists hasten to add, it is an independent branch, by
which they mean that language can, and should, be studied in its own terms,
without reference to psychological or neurological concepts and theories. On
this view, the role of linguistics is to provide hypotheses for other disciplines
to test. Some linguists are even more separatist: all too often, when research
on language acquisition or processing contradicts hypotheses developed on
the basis of purely linguistic data, they succumb to the temptation to ignore
it, justifying this decision by appealing to notions such as performance as
opposed to competence. This approach has resulted in a proliferation of lin-
guistic theories, many of which appeal to constructs which are extremely
implausible both psychologically and biologically, and an estrangement from
other disciplines studying the human capacity for language. It is becoming
increasingly clear to many linguists that in order to make further progress in
the study of language, we need to constrain linguistic theorising in some way.
Since human language resides in the human mind, the most promising source
of constraints on theories of language is properties of the human mind/brain.
This book is an attempt to show how linguistics can benet by incorporat-
ing insights from other disciplines. Part I, The Basic Specications, draws on
the relevant research in language processing, acquisition, and impairment,
comparative psychology, and genetics to develop a set of design specications
for language which should be the starting point of any linguistic theory that
purports to be psychologically realistic. The purpose here is not to engage in
debate or evaluate the research in these disciplines, but to concentrate on those
ndings which are relatively uncontroversial and which are of special impor-
tance to linguistic theory. For readers with little prior knowledge of these dis-
ciplines, this will be a useful introduction to the relevant psychological and
neurological issues.
These general specications are further developed in Part II, The Building
Blocks of Language, where they are applied to three central aspects of lin-
guistic organisation: lexical semantics, inectional morphology, and syntax.
Knowing how a system emerged often provides valuable clues about its
2 Language, mind and brain
organisation, so each of these areas is explored from a developmental perspec-

tive as well as from the point of view of the adult end state. Part II also
addresses several issues of central importance to linguistic theory: the status
of rules, the relationship between grammar and the lexicon, and the relation-
ship between innate structure and acquired knowledge. The nal chapter is
devoted to approaches to language which meet the specications developed
earlier, with particular prominence given to cognitive and construction
grammar.
Another recurrent theme throughout the book is the need to make linguis-
tics an empirical science. This means, above all, that linguistic theories should
be assessed on how well they explain the data and the empirical predictions
they make not on their formal elegance. It also means that, in order to reach
reliable conclusions, linguists need to make use of a wider range of data
sources.
Particularly important among these are examples of real usage collected in
large corpora of texts, both written and spoken. During the rst three decades
following the Chomskyan revolution, it was fashionable in certain circles to
assert that corpus data provides an inadequate foundation for linguistic
theory, for two main reasons. First, even the largest corpora necessarily
contain only an innitesimally small part of the possible sentences of any
given language, and may not contain any instances of the structures that are
of interest to the analyst. Worse still, corpus data is noisy, in that it often con-
tains ungrammatical sentences. Due to these limitations, it was proposed that
linguists should rely on grammaticality judgements rather than on collections
of actual texts.
As a matter of fact, grammaticality judgement data can also be very noisy.
Nave informants often reject grammatically well-formed sentences which are
semantically or pragmatically odd, or merely dicult to process, and accept
ungrammatical sentences in certain contexts. Because of this, many linguists
do not even bother to collect grammaticality judgements, preferring instead to
rely on their own intuitions, or those of a few colleagues. Such an approach,
however, is completely unacceptable if linguistics is to be regarded as a scien-
tic enterprise. A researchers own intuitions may be blunted after repeated
experience with ungrammatical sentences, or, worse still, contaminated by his
or her theoretical commitments. Therefore, while they are an excellent source
of hypotheses, they should, except in the clearest cases, be conrmed by a
systematic survey of native speakers judgements. This will also allow the
researcher to establish how much variation there is among speakers of the
same language, which could have far-reaching theoretical implications.
However, useful as they are, grammaticality judgements cannot replace
corpus data. Only the latter provides information about structures which
people actually use about language in its natural habitat, as it were and,
crucially, their relative frequencies, which often oers valuable clues about
how languages work. Furthermore, corpora sometimes yield unexpected
Introduction 3
information about systematic exceptions to generalisations proposed by lin-

guists a kind of data which we cannot aord to ignore. So, while the con-
cerns about corpus data enumerated earlier are valid, they do not entail that
linguists should not rely on this information source, but merely that they
should use other sources as well. Elicitation and grammaticality judgement
studies, for example, are a useful way of getting around the problem of acci-
dental gaps in corpora; and experiments in which the researcher manipulates
the language (e.g. by presenting speakers with a nonce word and asking them
to use it in a particular construction) yield valuable information about pro-
ductivity. In the following pages, we will use all of these data sources, since it
is only by combining information from a number of dierent sources that we
can hope to begin to understand the astonishingly complex phenomenon we
call language.
Part I
The Basic Specications
2 Language processing: speed and

exibility
1. The complexity of the task
Most people including many language professionals take language for

granted. We produce and interpret utterances so quickly and eortlessly that
its easy to overlook the enormous complexity of the task. It will be useful,
therefore, to begin our discussion by examining some of the diculties
involved in processing perfectly ordinary language. I will illustrate these prob-
lems using one simple sentence, picked more or less at random from a book
that happens to be lying on my desk as I am writing this David Lodges
Paradise News:
(1) Tess blew her nose on a paper tissue.
If you were to hear this sentence instead of reading it, what would reach
your ears would be a series of air vibrations; yet almost instantaneously you
would be able to work out its meaning: a woman picks up a piece of tissue,
covers her nostrils with it, and clears her nose by vigorously expelling air
through it. To arrive at this interpretation, you must have performed a number
of operations: identied the words; accessed their relevant senses; determined
the grammatical relationships between them; and integrated lexical, syntactic,
and contextual information with real-world knowledge to construct an inter-
pretation of the sentence in context. Each of these operations involves the
integration of multiple and only partially reliable cues.
Let us begin with the rst stage: word recognition. The major problems
here are variability of the speech signal and segmentation. The acoustic prop-
erties of speech segments vary according to the context in which they occur:
the /b/ sounds in blew, bet, bear and rob, for example, are phonetically quite
distinct. Furthermore, since the acoustic properties of speech depend on the
shape of the vocal tract and everybodys vocal tract is shaped slightly dier-
ently, what we perceive as the same sound is actually quite dierent when
produced by dierent speakers, or even by the same speaker on dierent occa-
sions. The segmentation problem is a consequence of the fact that speech is
continuous (speakers dont normally pause between words), and hence word
boundaries are not given in the acoustic input but must be inferred by the
listener.
Because of the lack of invariance and the segmentation problem, even this
low-level processing poses a considerable challenge, which is perhaps best
appreciated in the context of current developments in automatic speech rec-
ognition. Despite decades of intensive research, we are only beginning to see
computer applications that are reasonably accurate at identifying words in
continuous speech. Most of the existing programs are trained to recognise the
voice of a particular speaker in a particular acoustic environment, with little
background noise; and although their performance is good enough to make
them commercially viable, they still make many errors.
Once the listener has segmented the speech signal and identied the words,
he or she must access the relevant sense of each word and determine how they
combine to form constituents. Both of these operations are fraught with di-
culty, since most words are ambiguous, and there is often more than one way
of constructing a well-formed syntactic structure from a given string of words.
Let us begin with lexical ambiguity. The rst word in our sentence, Tess, is
a proper name, and hence has no conventional linguistic meaning apart from
specifying that the referent is female and probably human. However, the next
word, blew, is multiply ambiguous. As the past tense form of the verb blow,
blew could have any of the following meanings:
1. moved (of air particles), as in The wind blew in my face;

2. caused to move by blowing, as in The wind blew snow in my face;
3. sent out a stream of air from the mouth or nose, as in He blew (air) into
the balloon;
4. produced (e.g. smoke rings or bubbles) by blowing air out of ones mouth;
5. blew air into a wind instrument, as in She blew the trumpet;
6. (of a musical instrument) produced a sound: The trumpet blew outside;
7. burst, as in His tyre blew;
8. experienced a burst tyre while driving, as in He blew a tyre;
9. spent (money etc.), as in He blew his inheritance on a sports car;
10. wasted (an opportunity), as in She blew her chance;
11. burned out, as in The fuse blew;
12. exhaled vigorously through the nose in order to clear it.
(This is a selection of some of the most frequent meanings; any sizeable dic-
tionary will list many more.) The same phonological form, when spelt blue, is
also associated with the following meanings:
13. having a particular colour;

14. sad, as in Why are you feeling blue?;
15. pornographic, as in a blue movie.
Language processing: speed and flexibility 9
The next content word, the noun nose, can designate any of the following:
1. a part of the body;

2. the sense of smell, as in a dog with a good nose;
3. intuition, as in a reporter with a nose for a story;
4. the most forward part of an aircraft;
5. the open end of a pipe.
It can also have a range of verbal meanings:
6. (of an aircraft) to come down steeply;

7. to go forward carefully;
8. smell for;
9. search for;
10. pry into, as in nose into other peoples aairs.
The word paper also has several dierent senses:
1. substance manufactured from wood bre used for writing, drawing etc.;
2. newspaper;
3. a set of examination questions;
4. essay;
5. to cover with paper;
and the same is true of tissue, which can refer to any of the following:
1. woven fabric;
2. a mass of cells, as in muscular tissue;
3. a small piece of thin soft absorbent paper;
4. web, as in a tissue of lies.
In addition to these multiply ambiguous open-class words, our sentence

also has three function words, her, on, and a. These, too, can signal a variety
of meanings. Her can be either a possessive determiner (as in her nose) or the
objective form of the pronoun she (as in I gave her a pen). The preposition/par-
ticle on is one of the most highly polysemous words in English. Instead of
listing all the possible senses, which run into the hundreds, I will simply list a
few representative usages:
(2) a. All the books on politics are on this shelf.

b. She hung the picture on the wall.
c. She had a dog on a lead.
d. I wrote it on a piece of paper.
e. Have you ever worked on a farm?
f. The house is on the river.

g. We watched the boats on the river.
h. There werent many people on the train.
i. Ive never seen a smile on her face.
j. The meeting is on Thursday.
k. Does your car run on unleaded?
l. I heard it on Radio Four.
m. Shes on the phone.
n. The book was on special oer.
o. He read on.
p. Turn the lights on.
q. He is working on a new book.
r. Thats just not on!
Finally, the determiner a, in addition to various specialised uses which will not
be discussed here (but see Quirk et al. 1985: 26597), can be used to indicate
that the speaker wishes:
1. to refer to a specic member of the class designated by the following noun;

thus, one possible interpretation of Sally wishes to marry a millionaire is
that she wants to marry a particular millionaire;
2. to refer to any member of the class designated by the noun; thus Sally wants
to marry a millionaire could also mean that she is willing to marry anyone
with a seven-gure account (in other words, any old millionaire will do);
3. to refer to all the members of the class designated by the noun; thus A girl
should be realistic means that all girls should be realistic;
4. to quantify the number of objects referred to by the nouns; for example, in
a day or two means in one or two days.
To resolve these ambiguities, the listener must use grammatical, lexical, col-
locational, and pragmatic information. Let me illustrate this process using the
second word in our example, the phonological form /blu/. In the context of
our sentence, /blu/ is the past tense of the verb blow rather than the adjective
blue, since (1) there is no noun for it to modify and (2) it comes immediately
after the subject, and hence is likely to be the rst element of the verb phrase,
which is usually a verb or an auxiliary. The fact that the subject NP is animate
eectively eliminates senses (1), (6), (7) and (11), which normally co-occur
with subjects designating wind, a wind instrument, a tyre, and a fuse, respec-
tively. Senses (4), (5), (8), (9) and (10) are also ruled out for collocational
reasons: they normally take direct objects which designate a gaseous forma-
tion such as smoke rings (sense 4), a wind instrument (5), a tyre (8), money or
equivalent (9), or an opportunity (10). Our sentence could mean that she blew
into a nose-shaped piece of plastic, or that she sold her nose in order to buy a
paper tissue but such interpretations are extremely unlikely. Since noses
usually come attached to faces, sense (2) is ruled out on the basis of our knowl-
edge of the world (although the sentence could, in theory, mean that she pro-
pelled a plastic mould of a nose, or a real nose which had been severed from
someones face, by blowing until it landed on a paper tissue either by exhal-
ing air from the lungs or by using a device such as a bellows). The direct object
her nose is incompatible with sense (3), which is used in intransitive construc-
tions or with an NP designating the substance that comes out of the lungs such
as air or smoke. This leaves (12), the nose-clearing sense, as the preferred inter-
pretation of the phonological form /blu/ in the context of our example.
Let us now briey look at syntactic processing. English is a fairly rigid SVO
language, so the listener can safely assume that the preverbal NP (Tess) is the
subject. If we consider just lexical category information, the sequence blew her
nose could be interpreted either as a verb followed by a direct object, where
her is a possessive determiner in the object NP (analogous to ate his lunch) or
a verb with two objects (analogous to gave him lunch). However, her cannot be
the object of the verb, for two reasons. First, blow is monotransitive (it only
allows one object), so if her were its object, there would be no way of attach-
ing the NP nose to the verb phrase. Secondly, since nose is a singular count
noun, it must be preceded by a determiner. Therefore, the entire expression her
nose must be the object of blew.
The next four words, on a paper tissue, clearly constitute a prepositional
phrase. To interpret the sentence, the listener must determine how it combines
with the preceding words. There are three possibilities here: the PP could be a
complement to the verb, an adjunct, or a modier of the noun. In the rst case,
the sentence would have the same structure as Tess put the milk in the fridge.
This possibility is ruled out by the fact that blew is monotransitive, and hence
takes only one complement, the direct object. In the second case, on a paper
tissue would specify either the location or the direction of Tesss doings, and
the syntactic structure would parallel that for Tess read the book on the train.
Finally, the PP could postmodify the noun nose. Imagine that Tess has two
noses, and has put one of them on a paper tissue. In this context the sentence
could mean that she blew the nose which was on the paper tissue, and would
be similar in structure to Tess sold the house in Canada.
Our example is a very simple sentence, and hence the amount of syntactic
ambiguity that the reader or listener must cope with is fairly modest.1 However,
syntactic ambiguity becomes a major problem when the processor is con-
fronted with longer, more complex sentences characteristic of written dis-
course. It is well known from the natural language processing literature that the
number of candidate parses increases exponentially with sentence length.
Consider the following example, borrowed from Martin et al. (1987). Their
grammar assigns three structural analyses to the sentence List the sales of prod-
ucts in 1973 (the prepositional phrase in 1973 can modify the verb list, the
expression sales of products, or just the noun products). If we add a single word
to the sentence, making it List the sales of products produced in 1973, the
number of analyses increases to 10; and if we add with the products produced
in 1972 at the end of the last sentence, the number of alternative structural
descriptions (assuming the grammar developed by Martin et al.) is exactly 455.
Returning to Tess and her nose: to determine which of the two grammati-
cally acceptable readings identied above is the one that the speaker intended,
the listener must again use pragmatic information. Since noses are usually
attached to their owners and people dont normally have more than one, the
easiest way to specify which nose you mean is by identifying the owner by
means of a possessive modier; hence, a prepositional phrase after nose is
unlikely to be a postmodier. It must therefore be an adjunct specifying either
where Tess was when she blew her nose or where what she blew out of her nose
ended up. Stating that Tess was on a paper tissue is not a very good way of
locating her in space, since tissues are rather small in relation to a person, and
do not have a xed location. On the other hand, paper tissues are excellent for
blowing your nose on, so by far the most likely interpretation of the sentence
is that Tess used the tissue to clear her nose.
The listener must also decide whether her refers to Tess or to another female
not mentioned in the sentence. English grammar allows both possibilities, so
again, the correct interpretation depends on pragmatic knowledge. Since it is
impossible to clear another persons nose by exhaling air on a paper tissue,2 we
must conclude that her refers to Tess.
I have deliberately chosen to illustrate the complexity of the processing task
with a very simple (if somewhat insalubrious) sentence: one with no embed-
ding, no dislocated constituents, no long-distance dependencies. The example
illustrates rather nicely the amount of ambiguity that listeners must cope with,
as well as the fact that successful comprehension requires the integration of
lexical, syntactic and real-world knowledge. The discussion so far, however,
has completely ignored another important feature of ordinary language use
which introduces a whole new dimension of complexity: the fact that much of
the linguistic input that reaches our ears is degraded.
First, the acoustic signal itself is often distorted or partially inaudible
because of either background noise, or various modications that speakers
make in rapid speech, or both. In a now-classic study, Pollack and Pickett
(1964) spliced words out of a recorded conversation and asked speakers to
identify them. Only about 50 per cent of the words presented in isolation were
intelligible, but performance improved dramatically when the same words
were heard in a short phrase. By varying how the linguistic context was pre-
sented (visually or visually and auditorily), Pollack and Pickett were able to
show that the improvement was partially attributable to listeners adjusting to
the specic quality of the speakers voice and partially to the lexical and syn-
tactic constraints introduced by the additional material, a conclusion which
was corroborated by later research (see Massaro 1994; Samuel 1986). Thus, in
addition to all the complexities already discussed, listeners are also frequently
forced to guess what the speaker actually said.
Casual speech is also noisy in a metaphorical sense, in that it contains a

fair number of sentence fragments and false starts as well as some sentences
which are outright ungrammatical. This means that some of the grammatical
cues that language users can exploit when processing more carefully edited dis-
course (e.g. written texts) are unavailable or even misleading. We know very
little about how people process ill-formed sentences, although existing studies
suggest that morphosyntactic anomalies present surprisingly little diculty
(Braze et al. 2002; Jou and Harris 1991).
2. Speed
In spite of this complexity, humans process language extremely fast. In infor-

mal conversation people produce, on average, about 150 words per minute.
Furthermore, most conversations are reasonably contingent (people respond
to each others points) and proceed quite smoothly, with the next speaker
taking the oor within about 500 milliseconds (i.e., half a second) of the rst
speaker nishing his turn (Bull and Aylett 1998). Within this interval, the new
speaker must nish decoding the preceding utterance, decide on a relevant
response, construct her own message (or at least its beginning) and launch the
appropriate articulatory sequence. In fact, people often do this much faster. It
is not unusual for successive turns to overlap slightly (by about a syllable or
so). In such cases the second speaker has to do all her planning while her pre-
decessor is still executing his turn.
These observations about the speed of ordinary language processing are
conrmed by more controlled studies in laboratory conditions. One early
demonstration was a series of experiments by Marslen-Wilson and colleagues
(e.g. Marslen-Wilson 1973) in which participants were asked to shadow a
model, that is to say, repeat everything the model said as quickly as possible.
Some people are able to do this with only a quarter second latency in other
words, they are only about a syllable behind the model. Analysis of the per-
formance of such close shadowers suggests that they are not merely parrot-
ing a string of sounds, but processing the message at higher levels. For
example, they often correct errors (His knife tut the bread neatly is repeated
as His knife cut the bread neatly), and their own errors tend to be semantically
and syntactically appropriate (e.g. It was beginning to be light enough so I
could see . . . was shadowed as It was beginning to be light enough so that I
could see . . .; He had heard at the re brigade . . . became He had heard that
the re brigade . . .).
We can obtain more detailed information about the precise timing of the
various subprocesses from work on language comprehension. Studies of
spoken word recognition suggest that people can reliably identify words in iso-
lation after hearing the rst 30030 ms; and when presented in a sentential
context, many words can be identied after as little as the rst 200 ms
(Grosjean 1980; Marslen-Wilson 1987) well before the speaker has nished
articulating them!
Lexical ambiguity is also resolved very quickly. In fact, one of the most hotly
debated issues in psycholinguistics is whether the contextually irrelevant mean-
ings of ambiguous words are accessed at all, or whether access is selective from
the very beginning. This is a question of considerable theoretical signicance. If
comprehenders are able to use context to access only the relevant sense, we would
have to conclude that language processing is highly interactive. On the other
hand, initial activation of all senses is usually taken as supporting modular
accounts of processing, according to which language comprehension and pro-
duction is performed by a series of relatively autonomous components one spe-
cialising in word recognition, another in syntactic analysis, and so on.
Unfortunately, the large number of studies investigating the processing of
ambiguous words have produced mixed results: some studies support the selec-
tive access account, and others do not (see Simpson 1994 for a review). There is
an emerging consensus that the reason for these conicting ndings is that lexical
access is not an all-or-nothing phenomenon: the dierent senses of an ambig-
uous word are probably activated to varying degrees, depending on their relative
frequencies (with frequent meanings being inherently easier to activate) and the
strength of the contextual constraints. Thus, highly entrenched senses are di-
cult to suppress regardless of how inappropriate they may be in a given context;
but contextual factors seem to have a considerable eect on the activation of less
frequent senses.
However this debate is eventually resolved, the important point is that even
when several meanings are initially accessed, the contextually irrelevant ones
are deactivated very quickly. Tanenhaus et al. (1979) and Seidenberg et al.
(1982), for example, found that if an ambiguous word was presented immedi-
ately after a word which was semantically related to one of its senses, both
senses were primed that is to say, people were able to make lexical decisions
faster. If, however, the target word was presented after a short delay (about 200
ms), the facilitatory eect was found only when the prime was related to the
relevant sense of the target word, which suggests that the irrelevant sense had
already been eliminated at that point. This was true for noun/verb ambiguities
(He bought a watch/They decided to watch), which can be resolved using
morphosyntactic information, as well as noun/noun ambiguities (You should
have played the spade/Go to the store and buy a spade), which require the inte-
gration of linguistic and real-world knowledge.
These early studies have shown that semantic and pragmatic knowledge are
integrated with the incoming acoustic information extremely quickly. This was
conrmed by more recent studies which attempted to determine the point at
which participants establish reference by monitoring their eye movements as
they examined a visual display accompanying a linguistic stimulus. When
people are presented with a visual context for a spoken sentence, their eye
movements are closely synchronised with the unfolding spoken sentence. For
example, if they hear an instruction such as Put the apple on the towel while
viewing an array consisting of an apple, a towel and a few distractor objects,
they look at the apple on hearing apple; then, on hearing towel, they shift their
gaze to the towel. Thus, tracking a persons eye movements gives us a window
onto the mental processes underlying comprehension.
Using this methodology, Eberhard et al. (1995) were able to demonstrate
that when the visual array does not contain a distractor object with a similar-
sounding name, people are able to integrate acoustic, linguistic and visual
information and pick out the intended referent on average 55 ms before the
speaker has nished articulating the word. Altmann and Kamide (1999) went
one better and showed that in a suciently constraining context, listeners can
identify the probable referent before the speaker even begins the correspond-
ing word. In their experiment, people listened to sentences which contained
verbs with fairly specic selectional restrictions (e.g. The boy will eat the cake)
as well as more general verbs (e.g. The boy will move the cake). Both types of
stimulus were accompanied by a visual scene containing the objects men-
tioned in the sentence (in this case, a boy and a cake) as well as some distrac-
tors (e.g. a train set, a ball and a toy car). In this context, the verb eat is highly
predictive, since there is only one edible object in the scene, whereas move is
not. Listeners were able to use this information online and, in the eat condi-
tion, looked at the cake shortly after the oset of the verb.
Syntactic analysis and the integration of grammatical, semantic and con-
textual information also proceed at lightening pace so fast, in fact, that it is
very dicult to determine whether these are distinct processing stages, or
whether the processor immediately integrates all the information available.
Evidence for the former position comes from the so-called garden-path
eects. When confronted with structurally ambiguous input, readers and lis-
teners often have a strong preference for one analysis over the other (Ferreira
and Clifton 1986; Frazier 1987; Rayner et al. 1983). For example, on encoun-
tering the sentence fragment put the book on the shelf, people have a strong ten-
dency to interpret the prepositional phrase as a complement of the verb (a
phrase specifying where the book was put) rather than a modier of the noun
(specifying where the book was located before it was moved). Sentences in
which this expectation is violated (e.g. She put the book on the shelf in the box)
are more dicult to process, and the diculty manifests itself in longer
reading times in the disambiguating region (i.e. the second prepositional
phrase). The traditional interpretation of this nding is that the processor has
been lead up the garden path to an incorrect analysis, and additional process-
ing time is needed to reject it and compute a new one.3 Crucially, such eects
sometimes arise even when context or pragmatic considerations favour the dis-
preferred reading (Ferreira and Clifton 1986; Rayner et al. 1983). This has
been taken as supporting a modular view of language processing in which syn-
tactic analysis is informationally encapsulated and precedes semantic and
pragmatic processing. This conclusion, however, is controversial, since many
studies have found that semantic and pragmatic processing operate on the
same time scale as parsing, and, if the contextual bias is strong enough, can
even guide syntactic analysis (Altmann et al. 1998; Marslen-Wilson and Tyler
1987; Ni et al. 1996; Spivey et al. 2002; Trueswell et al. 1994). Moreover,
regardless of whether there is a purely syntactic stage, it is clear that other
types of processing are also performed very rapidly: the processing cost of a
garden path (the time required to reject the initial analysis and compute a new
one) is usually quite short: 100 ms or thereabouts.
The view that semantic and pragmatic processing begins immediately and
proceeds as quickly as parsing appears to be corroborated by research on
event-related brain potentials (ERPs). This involves measuring electrical
activity at the scalp while a person is engaged in some linguistic activity often
listening to sentences containing some kind of anomaly, since anomalies elicit
distinctive ERPs which can be linked to a specic point in the sentence. There
are three main ERP proles associated with language processing:
1. The left-anterior negativity (LAN) is a negative-amplitude wave in the left

anterior region which begins about 200 milliseconds after the onset of the
critical word and lasts for about 300 ms. Some researchers (e.g. Friederici
1995) argue it is associated with initial phrase-structure analysis based on
word category information; others (Kutas and King 1995) claim that it is a
reection of verbal working memory processes.
2. The N400 is a widely distributed bilateral negativity observed from about
200 ms until about 600 ms post-onset, with a peak at around 400 ms. It is
thought to be related to semantic processing.
3. The syntactic positive shift (P600) is a widely distributed positive wave
observed in the posterior parts of both hemispheres from about 500 to 800
ms post-onset, with a peak at around 600 ms. It is associated with a variety
of grammatical anomalies, including violations of phrase structure and
verb subcategorisation, subjectverb agreement, and reexiveantecedent
agreement.
Thus, ERP responses to semantic anomalies are observed at the same time
as (and possibly earlier than) responses to syntactic anomalies (Brown et al.
2000; Osterhout and Nicol 1999). Both begin while the word is still being
heard and are completed at the end of the word or shortly afterwards. In fact,
a latency of 200 ms post-onset suggests that the brain begins to react to the
anomaly even before there is enough acoustic information to identify the word
uniquely. This is conrmed by a study by Van Petten et al. (1999), who timed
the beginning of the N400 eect with respect to the isolation point the point
in the word at which most people are able to recognise it. Van Petten and col-
leagues identied the isolation point for a number of English words using the
gating technique. (This involves presenting people with very short initial frag-
ments of a word: e.g. the rst 50 ms, 100 ms, 150 ms and so on, until they have
identied the word correctly.) They then presented another group with sen-
tences which ended in a word which was either fairly predictable from the
context (e.g. He spilled some of the oil onto the engine because he didnt have a
funnel) or incongruous (He spilled some of the oil onto the engine because he
didnt have a tunnel) and measured ERPs for the nal word. They found, in line
with previous research, that the incongruous completions elicited a much
larger N400 than the congruous completions. The proles for the two condi-
tions began to diverge about 200 ms before the isolation point. This shows that
the participants in the experiment must have either anticipated likely comple-
tions for the sentences and matched them against the incoming signal, or used
the rst phonemes of the word to activate a cohort of candidate words and
compared their semantic representations with the semantic representation of
the sentence being processed. Whatever they did, the experiment shows that
the incoming acoustic information is immediately integrated with the repre-
sentation of the sentence at multiple levels.
3. Processing shortcuts
Thus, the enormously complex processing required to interpret an utterance

is carried out extremely quickly. To appreciate the signicance of this fact, it
is important to realise that human brains are fairly slow processors. Neurons
in the human brain typically re about 200 times per second. The clock speed
of most modern desktop computers is between 100 million and 1,000 million
cycles per second about a million times faster. Yet, while humans process
ordinary conversation quickly and eortlessly, computers run into serious
diculties at just about every level of processing.
So how do we do it? How is it possible for listeners to convert an acoustic
wave into a contextually interpreted utterance in a matter of seconds often
even before the speaker has nished articulating it? Or, looking at the issue
from the perspective of language production, how is it possible for speakers to
decide on the message, retrieve the appropriate lexical items, integrate them
into a grammatically coherent whole, and articulate the utterance in the same
time span?
The answer to this question comes in two parts. First, while individual
neurons are relatively slow, we have quite a lot of them somewhere in the
region of 100 billion, each connected to between 1,000 and 10,000 other
neurons. Consequently, a large number of computations are carried out in
parallel by dierent neurons or more realistically, by dierent assemblies of
neurons. This has profound (and as yet largely unexplored) implications for
theories of language processing.
Secondly, humans (and animals generally) are very good at using processing
shortcuts: simple heuristics, which, though not 100 per cent reliable, are good
enough for most situations (in contrast to algorithms, which are guaranteed to
provide the correct solution but are very demanding computationally). In the
following three subsections, we will explore some of these in relation to lan-
guage processing.
3.1 Prefabricated units

The simplest way of avoiding having to carry out complex computations is to
memorise the result and simply retrieve it from memory when required. In lin-
guistic terms, this means accumulating a large store of prefabricated chunks:
ready-made sentences and smaller units (phrases, inected words) which can
be combined into sentences with minimum eort.
It is incontrovertible that language users store some prefabricated units for
example, idioms such as spill the beans and social formulas such as How are
you? But to what extent does ordinary language use depend on such ready-
made chunks? To answer this question, we need to consider some basic facts
about human memory, the statistical structure of texts, and our daily experi-
ence of language.
3.1.1 We are able to store vast amounts of information

Our brains are superb information storage devices. There is a general consen-
sus among cognitive psychologists that, for all practical purposes, our long-
term memories can be considered to have innite storage capacity (Barsalou
1992: 148; Schneider and Bjorklund 1998: 469), and some researchers have
argued that encoding into memory is an obligatory, unavoidable consequence
of attention (Logan 1988: 493). We are also good at retrieving information
from memory when we need it, particularly if we have been repeatedly exposed
to it. What makes this possible is the fact that our memory is content address-
able: we can access memory representations virtually instantaneously on the
basis of just about any attribute. For example, on hearing a word we can
retrieve information about its meaning, grammatical properties, and the words
it typically collocates with. We are often able to do this on hearing just part of
a word: our brains immediately complete the pattern by lling in the missing
details. Similarly, we can activate a concept and retrieve the phonological and
grammatical properties of the corresponding word or words. If we can do this
for words, we can also do it for phrases and other complex units: in fact,
retrieval of complex units may be easier than retrieval of individual words of
comparable frequency, since the component units can act as retrieval cues
(provided that the complex unit has been analysed into its components and
that both are parts of the same associative network). For example, let us
assume that a speaker has stored the units open, door and open the door. The
last is considerably less entrenched than the former two (because it is less fre-
quent); however, it may still be relatively easy to retrieve because the speaker
can access the component units (open, door) which in turn will activate the
whole phrase.
3.1.2 The language we hear is full of recurrent word combinations

Erman and Warren (2000) examined a sample of spoken passages from the
LondonLund corpus and written passages from the LOB corpus. They found
that 59 per cent of the words in spoken discourse were parts of prefabs; in the
written sample, the proportion was slightly lower 52 per cent. This gure
probably underestimates the number of prefabs, since Erman and Warren
counted only word combinations which were in some way idiomatic. However,
there is no reason to assume that people dont also memorize fully composi-
tional units such as read a book, I am trying to . . ., or brothers and sisters.
To arrive at an upper bound for the frequency of prefabs in discourse, we
could simply count all the recurrent word combinations in a particular data
set. Such an analysis was undertaken by Altenberg (1998), who identied all
word combinations which occur at least twice in exactly the same form in the
LondonLund corpus. He found that the corpus, which comprises just under
half a million words, contained over 201,000 such combinations representing
over 68,000 dierent types. Together, these recurrent word combinations made
up over 80 per cent of the words in the corpus. Even this gure might be an
underestimate, however, since Altenberg counted only continuous combina-
tions, while many prefabs may contain gaps (e.g. for years, where the gap
can be lled by any numeral greater than 1, or I havent the idea, where
the gap has a more restricted set of llers: slightest, faintest, least, foggiest,
remotest: see Eeg-Olofsson and Altenberg 1993; Moon 1998).
3.1.3 People experience vast amounts of language in their daily lives

A frequency of two in a corpus of half a million words may not seem very
much. To put this gure in perspective, it will be useful to have an estimate of
how much language people experience in their daily lives. In normal adult con-
versation, about 150 words are spoken every minute (Huggins 1964; Maclay
and Osgood 1959). Much of what we hear on television and the radio news
reports, weather forecasts, commercials is considerably faster than this; and
skilled readers read 200400 words per minute. A rate of 150 words per minute
is equivalent to 9,000 words per hour. Assuming an eight-hour language day
that is to say, that most people spend about eight hours a day engaged in some
kind of linguistic activity (talking, reading, watching television, listening to the
radio, browsing the Internet etc.) they will hear or read about 72,000 words
in a typical day. This adds up to 504,000 words, or more than the entire
LondonLund corpus, in just a week. Thus, assuming the LondonLund
corpus to be representative, a frequency of two in the corpus equals about two
every week of ones life.4
3.1.4 People memorise prefabricated chunks

There are reasons for believing that people memorise multi-word chunks which
are much less frequent than this. Conventional multi-word units such as
idioms, proverbs, institutionalised metaphors, sayings and so on must be
learned precisely because they are conventional ways of expressing a particu-

lar idea. Yet corpus research suggests that most of them are fairly rare. In her
study of such phrasal lexemes Moon (1998) found that 70 per cent had a fre-
quency of less than 1 per million, and 40 per cent did not occur at all in her 18-
million-word corpus. Most of the expressions that Moon studied were drawn
from the Collins COBUILD dictionary, which is intended for learners of
English as a second language, and hence is quite selective in its coverage: it is
meant to be a guide to ordinary everyday English (Sinclair 1987: xix) and spe-
cically avoids any specialised or unusual expressions. Thus, we can plausibly
assume that most native speakers of English are familiar with the vast major-
ity of the phrasal lexemes that Moon studied including expressions such as
hang re, kick the bucket, one mans meat is another mans poison, speak for your-
self, which are among the 40 per cent that did not occur at all in the corpus. It
seems, therefore, that people have no trouble learning large numbers of idio-
matic multi-word units even if these are relatively infrequent. It doesnt follow
from this that people also memorise fully compositional units but it is not an
unreasonable conjecture, given that they have both the capacity and the oppor-
tunity. Note that in order to claim the opposite that people do not memorise
expressions which they are in principle capable of assembling themselves one
must assume that in addition to our memory and processing mechanisms we
are equipped with a screening device which either deletes redundant informa-
tion from memory or prevents it from being encoded in the rst place.
Nevertheless, most linguists (and many psycholinguists as well) take it for
granted that linguistic expressions are either computed or stored, and that if
a form can be computed, it is not stored. This is only an assumption, not sup-
ported by any evidence but it is a powerful one nevertheless, and as a result,
there has been relatively little work devoted to the storage of complex expres-
sions which are fully compositional. The one exception to this is work on reg-
ularly inected forms such as kissed or cats, much of it conducted in the
context of the debate on the dual mechanism theory of morphological pro-
cessing (see Chapter 8).
One of the tell-tale signs of memory storage is sensitivity to frequency:
memory representations of high-frequency items are more entrenched, and
hence are accessed more quickly and more reliably than low-frequency items.
Several studies have demonstrated frequency eects for regularly inected
forms, which suggests that at least some regular forms are stored. Stemberger
and MacWhinney (1986, 1988), for example, presented English-speaking
adults with regular verb stems and asked them to supply the past tense form
as quickly as possible. They found that their subjects made fewer errors on
high-frequency verbs, a result which is dicult to explain if one assumes that
all regular forms are constructed by applying a rule. Other studies suggest that
high-frequency regular forms are processed more quickly. In lexical decision
experiments, where participants have to decide as quickly as possible whether
a particular stimulus is a word or not, reaction times are shorter for high-
frequency forms, even when the frequency of the root is controlled for. This
nding has been reported for the Dutch third person singular (Schreuder et al.
1999), the Dutch plural (Baayen et al. 1997), and English noun morphology
(Sereno and Jongman 1997).5
Bod (2001) used a similar method to show people may also store simple sen-
tences. He presented people with high-frequency sentences (e.g. I like it), low-
frequency sentences (e.g. I keep it), and ill-formed strings consisting of the
same words (e.g. It I like). The high- and low-frequency sentences were iden-
tical in syntactic structure, were equally plausible and contained lexical items
of similar frequencies. The participants task was to decide as quickly as pos-
sible whether the word string was an English sentence or not; and they made
this decision faster for the high-frequency sentences.
Additional evidence suggesting that people store regularly inected forms
comes from studies which examined pronunciation. Word-nal stops in
certain consonant clusters are usually shorter and articulated less clearly than
in other positions, or they may be deleted altogether, particularly in rapid
speech: for example, band and guest may be pronounced [bn] and [s]
respectively. Such simplications are usually restricted to consonant clusters
which are part of the same morpheme: the nal clusters in banned and guessed
tend not to be simplied, in spite of the fact that these words are homopho-
nous with band and guest (New 1972). However, the nal [t] and [d] in high-
frequency past-tense forms behave like the corresponding segments in
monomorphemic words: they are often omitted in rapid speech, and, when
present, they tend to be pronounced less clearly (Bybee 1995, 2000), which
suggests that, like the monomorphemic forms, they are stored.
As indicated earlier, there is very little work investigating whether adults
store regular (i.e. non-idiomatic) combinations of words, although we do
know that children do, and that such units play an important role in language
acquisition (see Chapter 9). Bybee and Scheibman (1999) have demonstrated
that similar reductions occur in high-frequency word combinations: the form
dont in the expressions I dont know, I dont think, why dont you is phonetically
very dierent from the dont we nd in novel combinations, which, by the logic
outlined above, suggests that it is part of a stored chunk. There is also some
less direct evidence from language change. It is well known that frequently
occurring word combinations often fuse into single words. For example, in
English, going to, want to, have to are often pronounced gonna, wanna, hafta
(note, too, the conventionalised spelling); auxiliaries are often contracted
when they follow pronominal subjects (Im, weve, shell) and the negative par-
ticle not may attach to the preceding auxiliary (havent, doesnt, etc.). It is also
quite common for prepositions to fuse with the following article, as in the
German am (an + dem), zur (zu + der), aufs (auf + das) etc. For such phono-
logical reductions to occur, the participating words must have become single
chunks of structure in speakers minds, since they are not systematic but
restricted to specic combinations of words.
3.1.5 Prefabricated chunks facilitate language processing

We have seen that people store ready-made chunks of language probably
very large numbers of them.6 There is also evidence that the availability of
such stored units facilitates processing. I have already noted that high-
frequency regularly inected forms and sentences appear to be processed more
quickly and more accurately than low-frequency forms. Further relevant evi-
dence comes from research on idioms. People take longer to process the literal
meanings of conventional idiomatic expressions than the non-literal meanings
(Gibbs 1980; Mueller and Gibbs 1987); and idioms are processed faster than
comparable novel word combinations (Gibbs and Gonzales 1985; Peterson
and Burgess 1993). This explains why most people would never even consider
the various bizarre interpretations of Tess blew her nose on a paper tissue dis-
cussed at the beginning of this chapter: the availability of the semi-idiomatic
prefab X blow XPOSS-DET nose pre-empts the more compositional interpre-
tations.
There is also plenty of more informal and more ecologically valid evi-
dence. Bilinguals who use dierent languages in dierent situations for
example, one language at home and the other at work often avoid using a lan-
guage in the wrong setting (the home language at work or vice versa); and,
when forced to do so, can become surprisingly dysuent. All speakers, whether
bilingual or monolingual, tend to be more uent when discussing a familiar
topic. This is without doubt partly attributable to non-linguistic cognitive
demands, but also to the fact that they cannot rely on preconstructed phrases.
For similar reasons, as shown by Goldman-Eisler (1968), when people are
asked to retell the same story, they become more and more uent with each
retelling. They also tend to use the same expressions in the subsequent versions,
which indicates that the increase in uency is not due simply to speakers having
already decided which aspects to narrate and in what order. (In real life, you
can sometimes observe this phenomenon at dinner parties. It is usually fairly
easy to identify a fellow guests party piece, even if you have never heard it
before, because it is delivered much more uently than other anecdotes.)
Some particularly interesting evidence of the eect of using prefabricated
chunks on uency is presented by Kuiper (1996), who studied fast sport com-
mentators, auctioneers, and other individuals who can speak abnormally
quickly in certain situations. Their secret, Kuiper shows, is a large repertoire
of ready-made phrases, which they string together in a more or less linear
fashion. However, when not required to speak abnormally fast, these smooth
talkers switch to a dierent speech mode (colour commentary mode) which
is characterised by less stereotypical language and normal dysuencies.
3.2 Shallow processing

Since most prefabricated chunks are sentence fragments, in order to form a
complete utterance, speakers must usually piece together several prefabs or
combine prefabs with words. For example, in order to form an utterance such
as Why dont you stay at home tonight?, the speaker might integrate two pre-
constructed units (why dont you VP, stay at home) and the adverb tonight; in
order to construct our original example, Tess blew her nose on a paper tissue,
the speaker will need three chunks, X blow XPOSS-DET nose, on a N and
paper tissue, as well as the pronoun her and the proper noun Tess. Such piecing
together requires grammatical knowledge, since the component chunks must
be combined in a specic way if the result is to be grammatical. For example,
only units of a particular kind (viz., VPs with an untensed verb) can be
inserted into the slot after why dont you; the possessive determiner in the
second sentence must be coreferential with the subject; and so on. However,
the processing required to accomplish this is relatively shallow: the speaker
must attend to the relationships between units and the underspecied parts of
other units, but not to every detail of their internal structure.
There is growing evidence that the processing involved in ordinary language
comprehension is, in fact, fairly shallow: that is to say, people may integrate
the meanings of the most salient lexical items, some easily accessible grammat-
ical cues, and contextual information to arrive at a semantic interpretation
without fully parsing the sentence.
One of the earliest studies demonstrating this was conducted by Fillenbaum
(1974), who asked a group of students to paraphrase some sentences as accu-
rately as possible. Although they were specically instructed not to improve
the sentences or make them more sensible, the participants in the study often
normalised pragmatically anomalous sentences. For example, perverse
threats such as Clean up the mess or I wont report you were given a more plau-
sible paraphrase (e.g. If you dont clean up the mess I will report you) 54 per cent
of the time. The participants were then asked to reread each sentence and its
paraphrase in the presence of the experimenter and asked whether they could
see some shred of dierence between the two and even then, they failed to
spot the discrepancies 34 per cent of the time.
We cannot be sure what caused Fillenbaums subjects to give inaccurate par-
aphrases. They may have taken a pragmatic shortcut that is to say, arrived
at an interpretation without fully analysing the sentence; but it is also possible
that they interpreted the sentence correctly, but, since the interpretation didnt
make much sense, rejected it in favour of something more plausible. However,
the fact that some participants appeared to be unable to spot the discrepancy
between the meanings of the original sentences and their paraphrases of them
when they were specically invited to do so in the second phase of the experi-
ment does suggest that they took the pragmatic shortcut route.
Further support for this interpretation of Fillenbaums results is provided
by Ferreira (2003), who found that people often misinterpreted passive sen-
tences describing unlikely events: for example, when asked to name the
aected participant in The dog was bitten by a man, the participants said man
in over 20 per cent of all trials. However, they practically never made this
mistake with active sentences: The man bit the dog was virtually always inter-
preted correctly. So it is not the case that people have a tendency to correct
pragmatically odd sentences: normalisation seems to occur primarily with sen-
tences which are non-canonical, and hence fairly dicult to process.
In my own work (Dabrowska 1997), I have shown that shallow processing
can also occur in the absence of a pragmatic bias towards an incorrect inter-
pretation and is inversely correlated with educational achievement: that is to
say, the more formal schooling a person has had, the more likely they are to
make full use of the grammatical information available. The least educated
participants in the study were usually able to extract locally encoded informa-
tion, but had diculty in processing long-distance dependencies. For example,
when confronted with parasitic gap7 sentences such as It was her boss that
Louise persuaded that Margaret will interview, they were much better at iden-
tifying the interviewer (Margaret) than the interviewee (the boss). They were
also sometimes unable to establish hierarchical relationships between clauses
in multiply embedded sentences. Thus, the sentence Paul noticed that the fact
that the room was tidy surprised Shona was frequently interpreted to mean
Paul noticed that the room was tidy and this surprised Shona in other
words, the respondents appeared to extract simple clauses from the input (Paul
noticed something, the room was tidy, something surprised Shona) and then
attempted to integrate their meanings without making use of the grammati-
cal information available. Similar results have been reported by Chipere (2001,
2003).
Let us conclude the discussion with a particularly tricky example. Wason
and Reich (1979) found that nearly all the people they tested misinterpreted
the sentence No head injury is too trivial to ignore to mean No head injury
should be ignored, however trivial it may be. But according to the rules of
English grammar, the sentence cannot possibly mean this: it can only mean
the exact opposite. Consider the following:
(3) a. No child is too obnoxious to look after.

Every child should be looked after, no matter how obnoxious.
b. No question is too stupid to answer.
Every question should be answered, no matter how stupid.
c. No mistake is too trivial to correct.
Every mistake should be corrected, no matter how trivial.
Therefore:
(4) No head injury is too trivial to ignore.

Every head injury should be ignored, no matter how trivial.
Why is it that people systematically misunderstand the last sentence, but not
the other three? The problematic sentence is quite complex, both syntactically
and semantically. The noun phrase no head injury, which functions as the
theme of ignore, does not appear in its usual direct object position (as in We
ignored the head injury) but at the beginning of the sentence, as the subject of
the main verb, which makes it dicult to establish the relationship between
them. The sentence also contains three negatives: the particle no and two log-
ically negative expressions, too trivial (i.e. not serious enough) and ignore
(take no notice). Furthermore, it is very odd from a pragmatic point of view,
since it presupposes that serious matters should be ignored and trivial things
attended to (cf. This is too trivial to ignore), as well as enjoining the addressee
to ignore head injuries, which are widely known to be potentially dangerous.
It is not particularly surprising, then, that many people misinterpret the sen-
tence. What is interesting, however, is that they are so condent about their
erroneous interpretation and nd it extremely dicult to see the correct one,
even when they are told what it is.
3.3 Frequency
Fluent language users are remarkably sensitive to the statistical properties of
their language. This sensitivity manifests itself at all levels of linguistic organ-
isation, from phonemes and combinations of phonemes through to words and
grammatical constructions (Ellis 2000), and appears to be quite detailed.
People seem to know not just the relative frequencies of individual words and
word senses, but also how often they appear in specic constructions: for
example, that some transitive verbs (e.g. accuse) are much more likely to occur
in the passive construction than others (e.g. watch: see Trueswell 1996).
Similarly, when a verb allows more than one type of complement, people
know which kind of complement it typically takes: e.g. while both push and
move can occur in either a transitive or an intransitive construction, push
prefers the former and move the latter (MacDonald 1994); accept takes NP
complements more often than sentential complements, while suggest has the
opposite preference (Garnsey et al. 1997); imply is more likely than hope to
take an overt complementizer when it occurs with a sentential complement
(Trueswell et al. 1993); and so on (see MacDonald 1999 for further examples).
People are also sensitive to frequency information at more abstract levels.
For example, when confronted with an ambiguous expression such as
Someone stabbed the wife of the football star who was outside the house, where
the relative clause (who was outside the house) can modify either the upstairs
or the downstairs NP (the wife or the football star, respectively), English
speakers usually prefer the latter interpretation that is to say, in the absence
of contextual information supporting the other reading, they interpret the
sentence to mean that it was the football star, not the wife, who was outside
the house. However, speakers of many other languages, including Spanish,
normally choose the other interpretation when they hear or read the corre-
sponding sentence in their own language.8 This reects the statistics of the
language: corpus counts show that in English, a sequence consisting of a noun

followed by a prepositional phrase and a relative clause is usually associated
with the downstairs interpretation, whereas the opposite is true in Spanish
(Mitchell 1994: 398). A compelling piece of evidence demonstrating speakers
sensitivity to such factors is provided by an intervention study by Mitchell,
Cuetos and Corley (see Mitchell 1994: 398), who got two groups of Spanish-
speaking 7-year-olds to read stories which contained sentences with this kind
of ambiguity. In one set of stories, the relative clause modied the higher
noun, and in the other set, the lower noun. The children were matched in their
initial preference for one or the other interpretation; but when they were tested
two weeks later, their preferences had diverged.
Knowledge about the relative frequency of words and constructions mani-
fests itself primarily in ease of activation and processing times: at all levels of
linguistic organisation, high-frequency forms are inherently easier to activate
and hence are processed faster. Thus, when the input is ambiguous, people
consistently prefer the higher-frequency interpretation. By its very nature, pro-
babilistic information is not 100 per cent reliable. However, reliability
improves considerably when several dierent sources of probabilistic informa-
tion are combined (Redington and Chater 1998). The main advantages of
using probabilistic information is that it speeds up the processing of typical
cases that is to say, it speeds up processing most of the time and it also
helps to narrow down the search space when there are a vast number of alter-
native analyses, which is often the case with complex sentences found in
written texts (such as this one!). For this reason, there is a growing tendency
to incorporate probabilistic information into automatic parsers (cf. Bod 1998;
Briscoe 1994).
4. Implications for linguistic theory
Some linguistic expressions (e.g. irregular forms) are retrieved from the
memory and others (e.g. all novel expressions) must be computed. Still others,
such as frequently occurring word combinations and inected forms of
common words, can be either retrieved or computed. However, there is a
trade-o between storage and computation: a set of general, highly compact
rules does not require a large storage capacity, but poses severe demands on
the processing system; and conversely, the retrieval of prefabricated chunks is
computationally less demanding, but requires a large and ecient memory.
Linguists like general rules and principles. When choosing between compet-
ing accounts of the same phenomena, they have traditionally relied on the prin-
ciple of economy. General solutions are always preferred to more specic ones,
and it is assumed that any regularities that can be captured by means of general
rules and principles should be so captured. It is also held as a matter of course
that anything that is computable need not be, and hence is not, stored.
However, it is doubtful that the principle of economy is an appropriate stan-

dard by which to evaluate linguistic theories, if these are intended to be
psychologically realistic. Human brains are relatively slow processors, but
have enormous storage capacity. From a psychological point of view, there-
fore, retrieval from memory is the preferred strategy. People store vast
numbers of prefabricated units, and rely on them whenever possible even
when they are fully compositional and could also be assembled using rules. It
follows that linguistic knowledge is represented in a highly redundant manner,
and that the same expression may be processed in dierent ways on dierent
occasions.
It is important to bear in mind that most prefabricated units are fairly small:
between two and three words. Thus, in order to construct a complete utter-
ance, speakers must usually combine several prefabs, or a prefab and one or
more word-sized units. This requires linguistic knowledge since, if the result-
ing structure is to be well-formed, units of a particular kind must be combined
in a strictly dened way. Thus, a psychologically realistic grammar must
contain combinatorial mechanisms which can accommodate complex units as
well as simple units.
We have also seen that humans have very detailed knowledge about the fre-
quency of linguistic units, and are able to use this knowledge in resolving
ambiguities. In principle, it is possible that information about frequency is
stored separately from linguistic knowledge proper, as a kind of vast appen-
dix to the grammar. However, it is much more likely that our sensitivity to fre-
quency is simply a direct eect of the amount of experience we have with
various structures: frequent constructions are processed faster and with less
eort because people have had more opportunity to practise them. This would
suggest that our mental representation of language is shaped by language use,
and hence, it is counterproductive to draw a sharp distinction between com-
petence and performance.
Notes
1. Note, however, that there are a number of local ambiguities. As indicated

in the main text, her could be either a determiner introducing the noun
phrase functioning as the object of blow or the object itself: the ambiguity
is resolved when the processor encounters the next word, nose. Likewise, on
could be a preposition or an adverbial particle modifying blow (as in She
blew her nose on and on, meaning She continued blowing her nose).
Finally, on encountering paper, the processor could construe it as the object
of the proposition on rather than a modier of the following noun. Notice
that in each case the ambiguity is resolved at the very next word.
2. Not in the usual way, anyway. There are some rather disgusting possibil-
ities, but I will leave these to the reader to explore.
3. An alternative view is that the processor constructs several analyses in par-

allel, but the dispreferred ones take longer to assemble.
4. Of course, we have no way of knowing whether a particular expression
which occurred twice in the LondonLund corpus really has a frequency
of four per million. It is possible that it was only ever produced twice in the
entire history of the English language. Likewise, it is possible that many
fairly frequent expressions are not represented in the corpus at all after
all, compared to the whole body of texts in the language we call English,
or even the linguistic experience of a single native speaker, the corpus is
laughably small. However, assuming that the corpus contains a representa-
tive sample of spoken texts, we can be reasonably condent that the mean
frequency of the expressions that occur twice in the corpus is about four
per million.
5. Some studies (Pinker and Prince 1992; Ullman 1999) have failed to nd fre-
quency eects for regulars in English. This is probably a ceiling eect:
English past tense and plural inections are very simple, and hence com-
paratively easy to compute.
6. Pawley and Syder (1983: 192) estimate that speakers have hundreds of
thousands of such ready-made units in their repertoire.
7. A parasitic gap is a gap which has the same antecedent as another gap in
the sentence, and which could not occur if the sentence did not contain the
other gap: therefore, its presence is parasitic on that of the other gap. In
the example given in the text, the ordinary gap occurs after persuaded and
the parasitic gap after interview; the noun phrase the boss is the antecedent
for both.
8. I rst became aware of this ambiguity, and the fact that there may be lan-
guage-specic preferences in interpretation, when I was doing a Spanish
language course as an undergraduate. One day, during a discussion about
whether there was such a thing as true love, my teacher declared that El
nico amor verdadero es el amor del dinero, y el amor del perro, que se compra
(The only true love is the love of money, and the love of the dog, which
one buys). I interpreted the relative clause as modifying the word dog, and
dismissed the argument as completely irrelevant (It doesnt matter whether
you buy it or not, the important thing is that the dog loves you), whereas
my teacher obviously intended the other interpretation and insisted that
you cant talk about true love if it (love, not the dog) is bought. It wasnt
until after I left the class that I realised what he meant.
3 Language acquisition: robustness
Language acquisition, it is sometimes claimed, is highly sensitive to matura-

tional factors and surprisingly insensitive to environmental factors (Fodor
1983: 100; see also Crain and Lillo-Martin 1999; Gleitman 1981; Stromswold
2000). To be sure, language develops in all normal children exposed to lan-
guage, even in highly adverse circumstances: in neglected and abused children,
in cognitively impaired children, and in children who have limited access to
information from the external world due to sensory deprivation, i.e. deafness
or blindness. Furthermore, many researchers have been impressed by the fact
that children often go through similar developmental stages. In Stromswolds
words:
Within a given language, the course of language acquisition is remark-

ably uniform . . . Most children say their rst referential words at 9 to 15
months . . . and for the next 68 months, children typically acquire single
words fairly slowly until they have acquired approximately 50 words . . .
Once children have acquired 50 words, their vocabularies often increase
rapidly . . . At around 18 to 24 months, children learning morphologi-
cally impoverished languages such as English begin combining words to
form two-word utterances . . . Children acquiring such morphologically
impoverished languages gradually begin to use sentences longer than
two words; but for several months their speech often lacks phonetically
unstressed functional category morphemes such as determiners, auxil-
iary verbs, and verbal and nominal inectional endings . . . Gradually,
omissions become rarer until children are between three and four years
old, at which point the vast majority of English-speaking childrens
utterances are completely grammatical. (Stromswold 2000: 910)
Stromswold also notes some more specic commonalities. For example,

English-speaking children acquire the grammatical morphemes of their lan-
guage in roughly the same order: rst the progressive -ing, closely followed by
the locative prepositions on and in, then the plural -s, irregular past-tense
forms such as ate, slept and drunk, the possessive -s, the uncontractible copula
(as in Yes, she is), the articles, and regular past-tense inections. Still later come
the third-person -s and uncontractible auxiliary forms, and, nally, the
contractible copula and auxiliary (e.g. Shes tired, Hes coming) (see Brown
1973; de Villiers and de Villiers 1973). Auxiliaries and various complex con-
structions such as questions, Stromswold (1995) maintains, are also acquired
in the same order by nearly all children.
The universality and alleged uniformity of language acquisition are some-
times taken as evidence for an innate universal grammar which develops
according to a biologically determined timetable and does not depend on
other cognitive subsystems. As we shall see in this chapter, there are some
serious problems with this argument.
1. Individual dierences
First, we might observe that the remarkable uniformities that Stromswold

speaks of are not all that remarkable. We dont need to postulate a universal
grammar to explain why children produce single-word utterances before two-
word combinations, why the latter precede longer utterances, or why early
word combinations are generally telegraphic. What is more, even this rather
uninteresting progression is only a tendency: some childrens earliest utter-
ances are short phrases and sentences such as Look-a-that! or Whassat? which
are only later analysed into their component morphemes.
Second, many of the observed similarities can be easily explained by appeal-
ing to factors such as conceptual complexity, salience and frequency. For
example, the concepts of plurality and possession are more accessible to
young children than the purely grammatical concept of agreement; hence, the
plural and possessive suxes are learned earlier than the third-person -s; past-
tense forms of irregular verbs are generally acquired earlier than past-tense
forms of regular verbs because irregular verbs tend to be much more frequent
(nine of the ten most frequent English verbs are irregular); uncontractible
copula and auxiliary forms are acquired earlier than contractible forms of the
same items because they are phonetically more salient; the progressive inec-
tion and the prepositions in and on are acquired early because they designate
relatively simple concepts and are syllabic (and hence acoustically salient); and
so on.
Third, although there are many similarities between children acquiring the
same language, there are also vast individual dierences. Such dierences are
perhaps most obvious, and easiest to quantify, in lexical development. The
comprehension vocabularies of normally developing children of the same age
can dier tenfold or more (Bates, Dale and Thal 1995; Benedict 1979;
Goldeld and Reznick 1990), and there are also enormous dierences with
regard to what proportion of their expressive vocabularies children are able to
use in production: some childrens expressive vocabularies are nearly as large
Language acquisition: robustness 31
as their receptive vocabularies (i.e. they are able to produce almost all the
words they understand), while others may be able to understand more than
200 words, but still not produce any (Bates, Dale and Thal 1995). Children also
dier with regard to the kinds of words they learn in the initial stages of lexical
development. So-called referential children initially focus primarily on object
labels (that is to say, concrete nouns), while expressive children have more
varied vocabularies with more verbs and adjectives and a signicant propor-
tion of social routines and formulas such as stop it, I want it, dont do it (Nelson
1973, 1981). Finally, there are dierences in the pattern of growth. While many
children do go through the vocabulary spurt that Stromswold alludes to at
some point between 14 and 22 months, a signicant minority about a quarter
show a more gradual growth pattern with no spurt (Goldeld and Reznick
1990).
Grammatical development is also anything but uniform. Children may
begin to combine words as early as 14 months or as late as 25 months, and
show correspondingly large dierences in MLU (mean length of utterance)
and other global measures of grammatical development a year later (Bates,
Dale and Thal 1995). At 42 months, the dierences between the most and least
advanced normal children are equivalent to 306 months (Wells 1985). Some
children learn to inect words before they combine them into larger structures,
while others begin to combine words before they are able to use morphologi-
cal rules productively (Smoczynska 1985: 618; Thal et al. 1996). Some children
are extremely cautious learners, while others are happy to generalise on the
basis of fairly sparse evidence, which results in large dierences in error rates
(Maratsos 2000). Marked individual dierences have also been found in
almost every area of grammatical development where researchers have looked
for them, including word order, negation, case marking, the order of emer-
gence of grammatical morphemes and, yes, in the development of English
auxiliaries (Jones 1996; Richards 1990; Wells 1979), questions (de Villiers and
de Villiers 1985; Gullo 1981; Kuczaj and Maratsos 1983) and multi-clause sen-
tences (Huttenlocher et al. 2002; see also Lieven 1997 for a review).
Children also dier in acquisition styles, or the strategies they use to learn
to speak (Nelson 1981; Peters 1977; Peters and Menn 1993). Analytic or ref-
erential children begin with single words, which they articulate reasonably
clearly and consistently. Holistic or expressive children, on the other hand,
begin with larger units which have characteristic stress and intonation pat-
terns, but which are often mush-mouthed and may consist partly or even
entirely of ller syllables such as [dadada]. Peters (1977) argues that holistic
children attempt to approximate the overall shape of the target utterance
while analytic children isolate and produce single words. These dierent start-
ing points have a profound inuence on the course of language development,
since they determine how the child breaks into grammar. Analytic children
must learn how to put words together to form more complex units. They typ-
ically do this by rst learning how to combine content words, whereby they
produce telegraphic utterances such as Put ball sofa. Later on in development

they discover that dierent classes of content words require specic function
words and inections (nouns take determiners, verbs take auxiliaries and tense
inections etc.), and learn to supply these. Holistic children, in contrast, must
segment their rote-learned phrases and determine what role each unit plays in
the larger whole. Unlike analytic children, they sometimes produce grammat-
ical morphemes very early in acquisition, usually embedded in larger unana-
lysed or only partially analysed units. In other creative utterances, they may
use nonsense ller syllables often a schwa or a nasal consonant as place-
holders for grammatical morphemes. As the childs linguistic system develops,
these llers gradually acquire more phonetic substance and an adult-like dis-
tribution, and eventually evolve into function words of the target language
(Peters 2001; Peters and Menn 1993; see also Chapter 9). Thus, while both
groups of children may eventually converge on similar grammars, they get
there by following very dierent routes.1
There is evidence suggesting that the choice of language-learning strategy
is at least partially dependent on the childs environment (Hampson and
Nelson 1993; Lieven 1994, 1997; Pine 1994). The analytic style is typical for
rst-born children of middle-class parents, whose linguistic experience con-
sists mostly of one-on-one interactions with a parent. The expressive style, on
the other hand, is found more often in later-born children and children who
grow up in extended families, who engage more in polyadic interactions with
various family members. Occasionally, the same child may develop dierent
strategies when acquiring dierent languages. Elizabeth Batess daughter
Julia, for example, adopted a fairly analytic style in English, which she learned
mostly in dyadic contexts, and a more expressive style in Italian, learned pri-
marily in the context of large family gatherings (Bates et al. 1988).
It should be stressed that the existence of individual dierences is not neces-
sarily incompatible with a nativist view of language development. Dierences
in the rate of acquisition could be a result of dierent rates of maturation of
the language faculty; and many developmental asynchronies could be attrib-
uted to individual modules maturing at dierent times. Nativist theories can
also accommodate dierences in acquisition style, as long as they allow for
individual dierences in the genetic endowment for language, although this is
problematic to the extent that acquisition styles are linked to specic environ-
mental inuences and/or non-linguistic cognitive and aective factors. Two
points, however, should be clear from the above discussion. First, we cannot
argue for innateness on the basis of the alleged uniformity of language acqui-
sition, because language acquisition is not uniform. Secondly, large individual
dierences suggest that there is considerable exibility in the way that the
genetic program unfolds that is to say, the genetic program appears to be rel-
atively open.
2. The role of input
Language can develop in very inauspicious circumstances: in blind children,

who have limited access to contextual information that helps sighted children
guess the meanings of the adult utterances they hear; in cognitively impaired
children, who may be unable to make use of much of this information; and in
neglected children, who have limited access to linguistic data from which to con-
struct their mental grammars. Since all normal children develop language
except in situations of extreme deprivation, the argument goes, language acqui-
sition must be driven by internal processes, and does not depend on input except
as a trigger. In this section, we will critically examine the above argument.
It is undeniable that there are vast dierences in how much input children
get, and the quality of that input. In a study of childmother interaction,
Huttenlocher et al. (1991) found that the most talkative mothers in their
sample spoke up to 10 times as much per unit of time as the least talkative
mothers. Since parents also dier in the amount of time they spend with their
children, the actual dierences in the sheer quantity of language children hear
are probably even greater (see also Hart and Risley 1995, 1999). The usual
nativist take on such ndings is that they show that only a minimal amount of
input is necessary to trigger language development, since all normal children,
including those of taciturn or neglectful parents, do develop language.
Once again, however, the conclusion is premature. First, the amount of
input does matter, since children who hear less language develop language
more slowly (Hart and Risley 1995, 1999; Huttenlocher 1998). A recent study
of the linguistic abilities of children from socio-economically deprived back-
grounds found that half had moderate-to-severe language delays (Locke and
Ginsborg submitted; Locke et al. 2002). Since the children scored at age-
appropriate levels on a test of cognitive development, their poor performance
on linguistic tasks cannot be attributed to unfamiliarity with the testing situ-
ation or a general cognitive delay. The authors argue that the language delay
found in these children is a direct consequence of the simple fact that their
parents do not speak to them as much as middle-class parents do. Not surpris-
ingly, even more severe language delays are found in neglected children, chil-
dren of depressed and substance-abusing mothers, and institutionalised
children but, signicantly, not in abused children (Culp et al. 1991): the latter
presumably do get a good deal of language input, even if it does not occur in
a loving environment.
Quality matters, too. Certain maternal speech styles are associated with
faster language growth (Hampson and Nelson 1993; Murray et al. 1990). The
children of more directive mothers learn language more slowly (Della Corte
et al. 1983; Tomasello and Todd 1983), while children whose utterances are fre-
quently recast by their parents tend to make faster progress (Baker and Nelson
1984; Farrar 1990). There are robust correlations between childrens ability to
produce and understand multi-clausal sentences on the one hand and the
frequency of such sentences in the speech of their parents and, and, more
importantly, their teachers on the other hand (Huttenlocher et al. 2002). The
latter is theoretically more signicant, since, unlike the former, it cannot be
attributed to genetic factors. Many studies of twins have found that they are
often substantially delayed relative to other children (see Mogford 1993), pre-
sumably because they get less personalised input than singletons. But
perhaps the most convincing evidence for the relevance of input tailored to the
childs needs comes from studies of spoken language development in hearing
children of deaf parents. While such children are rarely completely cut o
from spoken language, they clearly get less of it than children born to hearing
parents. Even more importantly, they often get little personalised input from
adults, which has a profound inuence on their language development.
Consider, for example, Jim, a boy studied by Sachs et al. (1981). Jims
parents had very little oral language and communicated with each other in
sign, but they did not sign to him. He was exposed to English primarily
through television, which he watched frequently, but he also occasionally
played with hearing children. Sachs et al. began to study Jims linguistic devel-
opment when he was aged 3;9, just as he was beginning speech therapy. It was
immediately clear that Jims language was not just delayed, but qualitatively
dierent from that of normally developing children. Although he had a rea-
sonably good vocabulary (only slightly smaller than an average child of his
age), his comprehension was quite poor, and he also had severe articulation
problems. Perhaps as a result of these, he only produced very short utterances:
his mean length of utterance (MLU) at the beginning of the study was 2.2,
which would be appropriate for a child just over two. Unlike ordinary children
learning English, Jim made many word order errors, and often produced
decidedly un-English sentences such as those in (1).
(1) a. Not window two window.

b. My mommy my house a@sc play ball.
(?) I play ball at home with my mommy.
c. House a@sc chimney my house a@sc my chimney.
(?) My house has got a chimney.
d. This is how a@sc plane.
He also had particular diculty with inectional morphology and function

words, which he omitted 63 per cent of the time about twice as often as nor-
mally developing children with a comparable MLU.2
There are two important lessons to be learned from Jims case. First, mere
exposure to language does not automatically lead to acquisition, since he was
unable to learn English from the television.3 It seems that, in addition to
simply hearing a lot of language in context, children also need at least occa-
sional interaction with a competent language user who does the sorts of things
that parents do: repeat and rephrase utterances that the child did not under-
stand, and provide some sort of feedback on the childs production. This
could be in the form of expansion of the childs incomplete utterances, puz-
zlement at incomprehensible utterances, or simply a continuation of the topic
introduced by the child or compliance with a request, both of which would
indicate that the childs communicative intentions have been understood.
Secondly, Jims language shows that restricted input aects morphosyntactic
development as well as lexical development in fact, in his case the former was
clearly much more seriously aected than the latter. It is worth pointing out
that Jims case is not an isolated: Todd and Aitchison (1980) report similar
deviant constructions in the speech of another hearing child of deaf parents.
So both the quality and the sheer quantity of input do matter. On the other
hand, they do not seem to matter as much as one might expect, since children
who get only a tenth of the input addressed to the most priveleged children do
not take ten times longer to achieve the same level of competence. How can
we explain this apparent discrepancy?
It is important to bear in mind that most children get vast amounts of input.
Cameron-Faulkner et al. (in press) estimate that a typical toddler hears
between 5,000 and 7,000 utterances per day. This means that between the ages
of 1, when language acquisition may be said to begin in earnest, and 4, when
most children achieve productivity in the basic structures of their language,
they will have heard something of the order of 6,570,000 utterances, includ-
ing about a million WH questions, a million Y/N questions, 660,000 instances
of the full transitive construction (a subject followed by a verb followed by an
object), about 40,000 relative clauses, and almost 7,000 passives.4 So halving
the amount of input may have little eect on language learning simply because
most children get more input then they actually need: half of 6.6 million is still
a very large number.
The second point to bear in mind is that while reducing the amount of lan-
guage experience may have little eect on the linguistic development of some
children, it does aect others namely, children with low IQs, small short-
term phonological memory capacity, and sensory decits who appear to be
far more vulnerable to the eects of suboptimal input (Snow 1995). Thus,
quality and amount of input may help to explain an apparent paradox. There
are reports in the literature of children who developed language normally in
spite of very low IQ, small short-term phonological memory capacity, or other
impairments, which have led some researchers to conclude that language
development does not depend on such cognitive abilities. On the other hand,
speech therapy clinics are full of disabled children with severe language prob-
lems, which suggests that there are close links between linguistic and non-
linguistic development. We can reconcile these conicting facts by recognising
that non-linguistic cognition does matter but good input can compensate for
cognitive impairments (within limits); and, conversely, good language learn-
ers may be able to compensate for suboptimal input. However, when several
inauspicious circumstances conspire, language will suer.
This suggests that language may develop dierently in dierent circum-

stances. To investigate this possibility, we will consider one special population
blind children in more detail.
3. Language development in blind children
The linguistic development of blind children can be potentially very informa-

tive in the context of the debate on the role of innate factors in language acqui-
sition. If language acquisition involved merely setting the parameters in an
innate program, we would not expect to nd many dierences between sighted
and blind children: after all, both groups hear language, and both are
equipped with the same universal grammar, so both should follow the same
course of development. On the other hand, if language acquisition were
heavily dependent on experience, we might expect that blindness will have a
profound eect on the course of development, since much information that is
normally available to the child information about the visual properties of
objects and events in the environment, about what their interlocutors are
doing, what they are looking at, their gestures and facial expressions is inac-
cessible to blind learners.
How, then, does language develop in blind children? The short answer is
that it depends on the child. According to one study, about 40 per cent have
speech abnormalities of some kind (Elstner 1983). But most blind children
have no speech abnormalities, just a small delay; and some develop completely
normally (Gleitman 1981; Prez-Pereira and Conti-Ramsden 1999).
An additional factor which may account for these large dierences is the
fact that many blind children also suer from other handicaps. Deviant lan-
guage development seems to be the rule when the child has other disabilities
in addition to blindness or is growing up in an unfavourable social milieu.
According to Graham (1968, cited in Elstner 1983), almost half of all blind
children who are multiply handicapped are completely without language. The
incidence of language disorders in the group studied by Elstner, which
included a number of children who had other problems, was 76 per cent.
Furthermore, while most of the abnormalities in sighted children are phonetic
or phonological, the blind childrens problems tended to be lexical-semantic
and morphosyntactic; and there were more morphosyntactic abnormalities in
the speech of completely blind children than in children with partial vision.
The conclusion, then, is that blindness by itself does not cause language
diculties, although it often results in a small delay. However, when combined
with other problems, such as low IQ or an impoverished home environment,
it can lead to serious language problems.
But why are impairments which do not necessarily cause problems in other
children so disastrous in the blind, and why should blind children be more
likely to produce deviant language? The answer, I suggest, comes down to the
fact that blind children develop language somewhat dierently from seeing
children. Because many aspects of contextual information are not accessible
to them, blind children must make very good use of whatever information they
have got, and cognitive impairment may preclude them from doing so.
Furthermore, blind children have to be more proactive in obtaining informa-
tion about meaning. There are two ways in which learners can determine what
a new linguistic expression means: they can either observe competent speak-
ers use it and note as much as possible about the context, or they can experi-
ment that is to say, produce it and see what happens. Since blind children
cannot observe as much as sighted children, they must experiment more. The
result is that they often produce utterances which they clearly do not under-
stand. This meaningless language or verbalism of the blind is sometimes
regarded as an indication of deviant development, or even cognitive impair-
ment. However, it seems more illuminating to view it as a very good way of
obtaining clues about what linguistic expressions might mean.5
There is considerable evidence that blind children make more use of imita-
tion, repetition and verbal routines than seeing children, and tend to adopt a
holistic learning strategy (Dunlea 1989; Mulford 1988; Prez-Pereira 1994;
Urwin 1984). This suggests that they rely more heavily on distributional infor-
mation, presumably in order to compensate for the lack of visual information
(Prez-Pereira and Castro 1997). This language learning strategy, of course,
makes greater demands on the childs phonological memory. Thus, a short
phonological memory span, or poor auditory processing skills, will aect
blind children more than seeing children, and may be another reason for the
large individual dierences in the linguistic accomplishments of blind chil-
dren.
A particularly striking example of how a blind childs heavy reliance on
verbal routines may result in the production of very unusual structures is given
by Wilson and Peters (1988). Between the ages of 3;2 and 3;6, the child they
studied, Seth, produced a number of questions such as those given in (2) below
(Wilson and Peters 1988: 2513).
(2) a. What are we gonna go at [=to] Auntie and? [answer: Auntie and
Priya]
b. What did I get lost at the, Dad? [answer: at the store]
c. What are you cooking pan? [answer: pancakes]
Such utterances are peculiar by any standards; but they are particularly prob-
lematic for Principles and Parameters theory, since they apparently violate
universal constraints on movement. Sentence (2a) violates the coordinate
structure constraint, which states that items cannot be moved out of a coor-
dinate structure (or a more general constraint, subjacency, which states that
constituents cannot be moved across more than one bounding node: in
this case, NP and S). (2b) and (2c) are violations of the maximal projection
property of movement rules (the requirement that only maximal projections

can be moved to the specier position). In (2b), only the noun has been moved,
not the entire noun phrase, and in (2c), only the second part of the compound
noun.
This anomalous usage had its origins in an information-eliciting game that
Seths father used to play with him, in which the father produced an arma-
tive sentence with a blank and Seth had to ll in the blank, like that:
(3) a. [Getting ready for breakfast]

F: Put your . . .
S: Bib on.
b. [After breakfast]
F: What did you eat? Eggs and . . .
S: Mbacon. (Wilson and Peters 1988: 2623)
It seems that in the anomalous utterances in (2), Seth put such a sentence with
a blank in an interrogative structure.
These anomalous constructions in Seths speech were not, of course, a
direct result of blindness, but rather a consequence of his unusual linguistic
experience. They do, however, oer particularly compelling evidence that
unusual experience can result in unusual language production.
4. The robustness of language
We have seen that there are considerable individual dierences in the way that
children learn to use language. Some children rst produce single words
(mostly nouns) and later learn to combine them to form more complex expres-
sions; others begin with short phrases or sentences and extract words from
them; and still others probably the majority use both strategies. Some chil-
dren are rapid piecemeal learners and use a variety of grammatical construc-
tions in very early stages of development, before they have fully analysed
them; others look for more general patterns and consequently may appear to
be less advanced. Some children jump to conclusions on the basis of relatively
little data a risky strategy which results in high error rates; others are more
cautious and consequently make relatively few errors.
We have also seen that it is not true that language acquisition is insensitive
to environmental and non-linguistic cognitive factors. Environmental depri-
vation, sensory impairments and, as we shall see, cognitive decits all have an
eect on language development. However, unless several dierent factors con-
spire, children are usually able to overcome these adverse circumstances and
develop normal language. As we will see in the next two chapters, children with
very low IQs and children who have suered damage to the areas of the brain
which normally specialise in linguistic processing may also develop relatively
normal language, although the actual course of development may be quite

dierent from that found in normal children. And, of course, if a child is
exposed to two, three or even four languages, she will learn all of them, pro-
vided she gets enough input in every language.
The most striking feature of human language acquisition, then, is its robust-
ness. The language learning system is able to adapt exibly to a wide range of
very dierent circumstances. Clearly, this exibility is not innite; but within
limits, the language learning system applies whatever mental resources are
available to the (sometimes impoverished) input data to construct a functional
(though perhaps imperfect) system. This kind of exibility is incompatible
with an acquisition sequence whose exact course is predetermined genetically.
A rigid genetic program cannot make allowances for such diverse contingen-
cies of language development as exposure to more than one language,
extremely limited exposure to language, early brain damage, or lack of sup-
porting input from other modalities. The exibility of language development,
then, and the existence of a great deal of individual variation suggest that lan-
guages are learned, not acquired.
Notes
1. This is a somewhat idealised picture. Most children use a mixture of both

strategies.
2. It should be stressed that Jims problems are not be attributable to a lin-
guistic impairment, since he made excellent progress after intervention. His
MLU rose to 2.9 in just a month, and by age 6;11, he was performing above
age level on most measures of linguistic ability.
3. Note, too, that at the time that Sachs et al. began observing Jim, he had not
acquired sign language, in spite of the fact that he could observe his parents
signing on a regular basis. He did, however, make considerable progress in
sign later.
4. These estimates are based on frequency data given in Cameron-Faulkner
et al. (in press) for questions and transitives, Gordon and Chafetz (1990)
for passives, and Diessel and Tomasello (2000) for relatives.
5. I am not suggesting that blind children consciously adopt this strategy. On
most occasions, their goal is simply to get adult attention but they end up
getting a variety of clues about the meaning of their utterances.
4 Language in the brain
It has been known for well over a century that damage to the language areas
of the brain often has catastrophic eects on an individuals ability to produce
and understand language. Lesions in the third frontal convolution of the left
hemisphere, the so-called Brocas area, often lead to Brocas aphasia, which is
characterised by slow, eortful speech and diculties with function words and
inections (cf. example (1) below). Patients suering from this type of aphasia
are usually able to understand ordinary spoken language, although they may
struggle on formal tests of comprehension where they are deprived of contex-
tual information and must rely on grammatical cues alone. For example, they
may be unable to work out who chased whom given a semantically reversible
passive sentence such as The boy was chased by the girl, but will have no prob-
lems with The soup was eaten by the girl, where they can use a non-syntactic
strategy. Lesions to Wernickes area, i.e. the upper posterior part of the left
temporal lobe, in contrast, tend to lead to severe comprehension diculties in
the presence of uent but often deviant speech (cf. example (2)). Such patients
make frequent lexical errors, substituting semantically related but inappropri-
ate words (e.g. table for chair) or incorrect sounds (e.g. bekan for began), or, in
more severe cases, unidentiable phonological forms such as taenz.
(1) [Retelling the story of the fox and the crow] King . . . Singing . . .
Singing loud . . . Meat. Perfect! (Goodglass 2001: 64)
(2) [Response to a therapists question about why he is in hospital]
Whats wrong with me because I . . . was myself until the taenz took
something about the time between me and my regular time in that
time and they took the time in that time here and thats when the the
time took around here and saw me around in it its started with me no
time and then I bekan work of nothing else thats the way the doctor
nd me that way. (Obler and Gjerlow 1999: 43)
The frequent occurrence of such impairments following damage to Brocas

and Wernickes areas indicates that these regions are crucially involved in lan-
guage processing. It has also led some researchers to speculate that linguistic
Language in the brain 41
knowledge is actually embodied in the wiring of these areas of the brain: spe-
cically, that Brocas area is the seat of grammatical knowledge while lexical
knowledge is at least partially represented in Wernickes area. As we shall see,
however, such claims are rather problematic for a number of reasons.
1. The localisation issue
The rst problem with the strict localisation view is the fact that the relation-
ship between the lesion site and the nature of the resulting impairment is far
from straightforward. Language impairment sometimes occurs after brain
damage outside the classical language areas (Dronkers et al. 2000), even in
right-handed monolingual patients with no known prior neurological condi-
tion which might have caused a reorganisation of language function in the
brain. If fact, aphasic-like symptoms may occur after damage to subcortical
areas like the basal ganglia and the thalamus (Basso et al. 1987; Brown 1975;
Crosson et al. 1986; Dmonet et al. 1991; Lieberman 2002). This suggests that
many regions of the brain outside the classical language areas are involved in
language processing a supposition conrmed by a growing number of brain
imaging studies (Dmonet et al. 1992, 1993; Mazziotta and Metter 1988;
Posner et al. 1988; Warburton et al. 1996; Wise et al. 1991; Zatorre et al. 1992).
Conversely, damage to the language areas does not always lead to language
impairment, or to the predicted type of impairment. Dronkers et al. (2000)
describe patients with damage to Brocas area who are not agrammatic, and
patients with lesions to Wernickes area who do not have Wernickes aphasia.
Furthermore, persistent full-blown aphasias are rarely found in patients with
damage to Brocas or Wernickes area alone: when the lesion is conned to just
the classical language areas, patients usually recover (Dronkers et al. 2000;
Lieberman 2002).
Recovery after damage to the language areas of the brain is particularly
good in children. As a rule, children recover quite quickly from the eects of
the lesion, and, if it occurs before the onset of language, the expectation is that
they will make a more or less full recovery that is to say, within a few years,
their language abilities will be within the normal range, though usually slightly
below normal controls (Aram 1998; Bates 1999; Bates et al. 1997). Lesions sus-
tained in middle and late childhood typically do leave some decits, although
these are more likely to be apparent in the childrens writing and schoolwork
than in ordinary conversation (Bishop 1993; Martins and Ferro 1992, 1993;
van Hout 1991).
It is also worth noting that the aphasic symptoms observed in children who
have suered brain damage are very dierent from those found in adults.
Wernickes (uent) aphasia is extremely rare in children, even after damage to
Wernickes area, as is the telegraphic speech characteristic of Brocas aphasia.
Child aphasics typically dont talk very much, and when they do, they use
simpler structures but they rarely make errors. Furthermore, children are
likely to have a mild aphasia after just about any kind of brain damage, includ-
ing damage to the right hemisphere and the non-linguistic parts of the left
hemisphere (Aram 1998; Gardner 1974). This suggests that the localisation of
linguistic functions is much more widely distributed in children than in adults,
which doesnt t in very well with the idea that the basic blueprint for grammar
is genetically prespecied and embodied in the wiring of Brocas area.
But the most dramatic illustration of the plasticity of linguistic organisa-
tion comes from studies of individuals who lack the left hemisphere alto-
gether. Some cases of catastrophic childhood epilepsy can be cured or at least
alleviated either by removing the entire hemisphere in which epileptic attacks
begin (hemispherectomy) or, more commonly, by stripping it of its cortex
(hemidecortication). Not surprisingly, children who have undergone such
drastic surgery suer from a variety of decits, including linguistic decits.
However, the nature of their impairment is often very dierent from what one
might expect, given what happens in brain-damaged adults.
It is important to bear in mind that the hemispherectomised population is
very heterogeneous: patients dier in the actual condition which lead to the
operation, the age at which it appeared, the age at which they were operated
on, the success of the operation in controlling seizures, and the state of the
remaining hemisphere. All of these factors, and probably many more, have an
eect on how well language develops after surgery. Consequently, the nal
outcome after hemispherectomy is extremely variable. However, given the
special status of the left hemisphere in language processing, we would expect
its removal to have a dramatic and devastating eect on an individuals lan-
guage abilities, while removal of the right hemisphere should normally spare
language (although it might impair the ability to use prosody and to interpret
gurative language, both of which are known to require right-hemisphere
involvement). This expectation, however, is not corroborated by the available
evidence.
A recent survey of 49 post-hemispherectomy patients (Curtiss and de
Bode 1998) found no signicant relationship between side of lesion and lan-
guage outcome. In fact, nearly a third of the right hemispherectomised
patients had no productive language at all, compared to 16.7 per cent of the
left hemispherectomies, which suggests that the right hemisphere plays an
important role in language acquisition. Even more surprisingly, seven of the
thirty left-hemispherectomised patients that is to say, over 23 per cent had
normal or nearly normal language.
Curtiss and de Bode acknowledge that their measure of linguistic ability,
the Spoken Language Rank, is fairly crude, so it is possible that a more sen-
sitive test of morphosyntactic abilities would reveal subtle decits in the left-
hemispherectomised group. Such decits were found by several researchers,
including Dennis and Kohn (1975), Dennis and Whitaker (1976), Dennis
(1980), and Vargha-Khadem and Polkey (1992). Dennis and Kohn (1975), for
example, used the Active-Passive-Negative test, in which the patient is pre-

sented with two pictures (e.g. a picture of a girl pushing a boy and a boy
pushing a girl) and a sentence (e.g. The boy is being pushed by the girl) and
asked to point to the picture that matches the sentence. The sentences vary in
voice (active v. passive) and polarity (armative v. negative). Dennis and
Kohn found that both groups performed well on active sentences, but the left-
hemidecorticate group had more problems with passive and especially nega-
tive passive sentences, from which they conclude that the right hemisphere is
not as procient at linguistic processing as the left hemisphere. This conclu-
sion, however, has not gone unchallenged. Bishop (1983) argues that perfor-
mance on passive negatives is low in children and in individuals with low IQ,
and that the left-hemidecorticates performance was not in fact signicantly
dierent from that of age- and IQ-matched controls; and Ogden (1988) found
that the decits in left-hemidecorticates performance on passive-negative sen-
tences disappeared when the test sentences were presented in writing rather
than orally.
Thus, both left- and right-hemidecorticates typically have language prob-
lems, although it may be the case that the left-hemidecorticates are less pro-
cient in the more complex aspects of morphosyntax. It must be stressed that
the decits are detectable only through specialised tests: the spontaneous
speech of the more procient left-hemidecorticated patients is uent and
grammatical. It is clear, then, that the isolated right hemisphere is capable of
sustaining a great deal of syntactic processing.
2. Preservation of grammatical knowledge in Brocas aphasia
There is also growing evidence that agrammatics, and aphasic patients in

general, may not have lost their linguistic knowledge. Aphasic patients linguis-
tic performance is often highly variable, and may improve or deteriorate con-
siderably from day to day, or at certain times of the day (Goodglass and Kaplan
1972: 12; Linebarger 1989: 199). It is also very dependent on the nature of the
task: patients often perform better on more structured tasks. For example, they
might be able to complete a passive sentence begun by a therapist even if they
are unable to construct one themselves. Furthermore, agrammatic production
does not necessarily co-occur with agrammatic comprehension: some patients
can understand grammatical distinctions conveyed by word order and function
words but are unable to use the same constructions in their own speech, while
others have relatively normal speech but cannot use grammatical information
in comprehension (Blackwell and Bates 1995; Linebarger et al. 1983). To make
things even more complicated, some cases of agrammatism are modality-
specic: the symptoms may be manifest only in the patients speech, or only in
writing. All these facts suggest that agrammatic patients (or at least those
agrammatic patients who show such uneven proles) have not lost their
grammatical knowledge, but have diculty accessing it in other words, that

their decit is not one of competence, but of performance.
There is also evidence that even truly agrammatic patients show some sen-
sitivity to grammatical structure. In one study (Wayland et al. 1996) patients
were asked to listen to sentences and press a button when they heard a partic-
ular word. The researchers found that agrammatic participants reaction times
were slower when the sentence was ungrammatical (e.g. when a verb occurred
in a syntactic context appropriate for a noun or vice versa). These results
suggest that the agrammatics were sensitive to the morphological cues signal-
ling grammatical category even though they did not produce the relevant mor-
phemes in their own speech. Even more intriguingly, there is now a whole raft
of studies showing that agrammatic patients often perform surprisingly well
on grammaticality judgement tasks (see Linebarger 1989, 1990; Linebarger et
al. 1983; Wulfeck and Bates 1991; Wulfeck et al. 1991). Linebarger et al.
(1983), for example, asked agrammatics to evaluate the acceptability of sen-
tences such as those in (3)(5) below, and found that they gave the correct
response about 80 per cent of the time a level of performance which is far
from perfect, but nevertheless well above chance.
(3) a. Did the old man enjoy the show?

b. *Did the old man enjoying the show?
(4) a. I want you to go to the store now.
b. I hope you will go to the store now.
c. *I want you will go to the store now.
d. *I hope you to go to the store now.
(5) a. Frank was expected to get the job.
b. Which job did you expect Frank to get?
c. *This job was expected Frank to get .
It is worth noting that many of the ungrammatical sentences used in the study
were locally grammatical, so the relatively high level of performance could not
have been achieved if participants carried out only a supercial analysis. As
the authors point out:
At least the following kinds of syntactic ability are required by our task:
an awareness of strict subcategorisation frames, considerable sensitivity
to the function vocabulary, the ability to compute hierarchical phrase
structure, and the ability to handle discontinuous syntactic dependen-
cies. It appears that the task could not be performed successfully by
using general semantic knowledge or an appreciation of transitional
probabilities and/or prosodic patterns associated with particular lexical
items. In general, any appeal to template matching as an alternative
strategy for performing this task seems to us to require templates which
are in fact generalizations over structures of considerable syntactic spec-

ication. (Linebarger et al. 1983: 386)
Yet another source of evidence suggesting that aphasic disorders may

involve lack of access to grammatical knowledge rather than lack of knowl-
edge as such comes from studies of bilingual aphasics. Most bilingual speak-
ers suering from aphasia are impaired in both of their languages, but in a
signicant minority of cases, the speakers languages may be dierentially
aected. In some patients, one language is severely impaired while perfor-
mance in the other remains relatively good. Others show unexpected patterns
of recovery: for example, the two languages may recover in succession, or the
patient may improve in one language but not in the other, or one language may
recover while the other deteriorates (see Paradis 1977, 1998). Occasionally
patients may show dierent symptoms in dierent languages: for example, a
patient may behave like a Brocas aphasic in one language and like a
Wernickes in another (Albert and Obler 1978; Silverberg and Gordon 1979).
But perhaps the most exotic case of linguistic impairment following brain
damage is a condition known as alternate antagonism with paradoxical trans-
lation described by Paradis et al. (1982). Their two bilingual patients could
use both of their languages spontaneously, but not at the same time (alternate
antagonism). Moreover, they were able to translate into the language that was
not available for spontaneous production, but not into the other language
(paradoxical translation). For example, one of their patients, a French/Arabic
bilingual referred to as AD, was on some occasions able to speak French but
not Arabic and to translate into Arabic but not into French; on other occa-
sions, she showed the opposite pattern. Clearly, AD could not have lost either
of her languages, although she was not able to access them at will.
The ndings reviewed above suggest that the linguistic impairments experi-
enced by many aphasic patients are attributable to processing diculties, not
to loss of linguistic knowledge (although global aphasia probably does involve
loss of knowledge). This interpretation appears to be supported by several
studies which showed that aphasic-like symptoms can be induced in normal
people under cognitive overload or when the stimuli are partially masked by
noise (Blackwell and Bates 1995; Dick et al. 2001; Miyake et al. 1994; Tesak
1994). Miyake et al., for example, got their subjects to read sentences pre-
sented to them word by word at a very fast rate. This method of presentation
signicantly reduced the subjects ability to use grammatical cues in interpret-
ing relatively complex grammatical structures such as relatives, but had little
eect on simpler sentences. In the Blackwell and Bates study, a group of
undergraduates had to keep a series of digits in memory while performing a
grammaticality judgement task. Their performance, though superior to that
of aphasics, was clearly impaired relative to no-digit controls. Interestingly,
agreement errors were especially dicult to detect in such conditions. Since
aphasics also have particular diculty with agreement morphology, the study
oers strong evidence in favour of the claim that the selective vulnerability of
grammatical morphemes in aphasia may be at least partly attributable to
global diminution in cognitive resources.
Most of the work reviewed so far is based on aphasia in speakers of English
and other typologically similar languages. In recent years, however, it has
become increasingly clear that brain damage may aect speakers of dierent
languages in dierent ways and consequently, that some of the conclusions
about the organisation of language in the brain arrived at on the basis of data
from the familiar European languages may be in need of revision. Two main
ndings are relevant in this connection. First, it appears that users of typolog-
ically very dierent languages may suer aphasia after lesions which do not
cause linguistic problems in speakers of English. The traditional literature on
aphasia, for example, is virtually unanimous in asserting that in the vast
majority of cases, language impairment occurs after damage to the dominant
hemisphere i.e. the left hemisphere in right-handers and the right hemisphere
in most left-handers. Damage to the non-dominant hemisphere may aect a
patients pragmatic skills, but it rarely results in serious decits in the core lin-
guistic areas of lexis and grammar. However, this does not appear to be the
case in Chinese. Hu et al. (1990) reported that in their sample of Chinese right-
handed aphasics, about 30 per cent had suered damage to the right (i.e. non-
dominant) hemisphere. This suggests that Chinese speakers, or at least a large
proportion of Chinese speakers, make much greater use of the right
hemisphere when speaking than do speakers of Indo-European languages.
Researchers working on sign language have also found that linguistic decits
in signers can occur following damage to the right hemisphere and some non-
linguistic regions of the left hemisphere as well as Brocas and Wernickes
areas, suggesting that the areas of the brain involved in processing sign lan-
guage only partly overlap with those used by spoken languages (Corina 1998).
These ndings have been conrmed by imaging studies (Corina 1996; Neville
and Bavelier 2000). Thus, it appears that experience with dierent languages
might cause speakers to recruit additional regions of the brain in addition to
the classical language areas.
Secondly, it has been shown that the traditional aphasias, and Brocas
(non-uent) aphasia in particular, look very dierent in typologically dier-
ent languages. One of the hallmarks of Brocas aphasia in English-speaking
patients, apart from dysuency and other problems with articulation, is
agrammatism, which is often manifested as telegraphic speech: the patient is
able to produce nouns and some other content words, but has severe diculty
with function words and grammatical endings, resulting in speech output rem-
iniscent of the style we nd in telegrams. However, non-uent aphasics in
highly inected languages such as Polish, Serbo-Croatian, German and
Italian, and agglutinative languages such as Turkish, Finnish and Hungarian,
are not agrammatic that is to say, they do not omit inections, though they
often use them inappropriately (Bates et al. 1991; Menn and Obler 1990;
Slobin 1991). Slobin (1991), for example, found that inectional morphology
in Turkish Brocas aphasics was remarkably well preserved . . . with hardy a
trace of telegraphic speech (1991: 150). His subjects tended to provide the
right grammatical morphemes where required, although compared to normal
controls, the range of axes in their speech was quite restricted similar to
the range of forms available to Turkish 3-year-olds. Slobin also notes that in
cases of word-nding diculty, patients often used pronouns instead of
nouns, and these were correctly inected. Even the most non-uent patient,
whose speech consisted mostly of single word utterances, provided the appro-
priate case inections on nouns.
Thus, the results from highly inected languages are at odds with those
obtained from English-speaking patients. The agrammatic patients described
by Slobin and Bates et al. had suered similar lesions to agrammatic English-
speaking aphasics, yet they had clearly not lost all grammatical knowledge. It
seems safe to conclude, then, that grammatical knowledge does not reside in
the pattern of connections in Brocas area.
3. The co-occurrence of lexical and grammatical decits
In popular treatments of language disorders, particularly those intended for

consumption by linguistics students, Brocas aphasia is often treated as a pri-
marily grammatical impairment, while Wernickes aphasia is presented as a
lexical decit. Such a characterisation, though it ts in very well with the tra-
ditional division of a speakers linguistic knowledge into a grammatical and
a lexical component, is very misleading.
First, it is now well known that Wernickes aphasics, in addition to their
obvious lexical problems, also have grammatical decits. Although they do
not omit function words and grammatical inections in production, they tend
to use them inappropriately and often produce ill-formed structures. This is
particularly obvious in highly inected languages such as Polish, Czech or
German, where substitution of an inappropriate grammatical morpheme is
likely to result in a sentence which is ungrammatical rather than merely
semantically anomalous (Pick 1973 [1913]; Ulatowska et al. 2001). They also
perform poorly on grammatical comprehension tasks (Kertesz and Osmn-
Sgi 2001; MacWhinney et al. 1991) and grammaticality judgement tasks
(Grodzinsky and Finkel 1998).
Conversely, the agrammatic performance of Brocas aphasics is virtually
always accompanied by word-nding diculties (particularly with verbs and
other relational words) and other disturbances of lexical processing such as
reduced semantic priming (Dick et al. 2001). Indeed, it appears that lexical
decits of some description accompany all forms of aphasia (Dick et al. 2001;
Marshall 1986), although their severity does not necessarily correlate with the
severity of the patients other decits. Furthermore, as pointed out by
Elizabeth Bates and her collaborators (Bates and Goodman 1997, 1999; Dick
et al. 2001), there are some rather striking parallels between a persons lexical
and grammatical diculties, which can be observed in a wide range of de-
cits, including developmental disorders, aphasia, and the linguistic impair-
ments which accompany Alzheimers dementia and normal ageing. Thus,
individuals who omit function words also tend to omit lexical items rather
than substitute inappropriate ones. This pattern is found in Brocas aphasia as
well as in other non-uent decits such as Down syndrome and some forms
of specic language impairment. In contrast, lexical substitution errors
(semantic paraphasias) and the substitution of inappropriate grammatical
functors (paragrammatism) are the hallmarks of Wernickes aphasia, and are
also quite characteristic of the speech of individuals suering from another
uent disorder, Williamss syndrome. Finally, the excessive use of pronouns
and semantically light forms such as thing, make and do in anomia and
Alzheimers disease (and, to a lesser extent, in normal ageing) is accompanied
by a similar reduction of the grammatical repertoire: individuals belonging to
these groups tend to avoid complex syntactic structures such as passives
(Bates, Harris et al. 1995). The co-occurrence of lexical and grammatical de-
cits in all these groups and the similarity in the symptoms that they demon-
strate across linguistic domains strongly suggest a common underlying
mechanism.
4. The resilience of language
It is undeniable that some regions of the brain are more involved in linguistic,
and specically grammatical, processing than others. However, the strongest
version of the anatomical specialisation hypothesis that grammar resides in
the pattern of connections in Brocas area is clearly false. As we have seen,
there is considerable evidence that individuals who have suered lesions to
Brocas area do not lose their grammatical knowledge, but are simply unable
to access it at will. Furthermore, the most entrenched grammatical patterns,
such as basic word order or case inections in morphologically rich languages,
generally do remain accessible. This suggests that linguistic knowledge is rep-
resented in a redundant manner in various regions of the brain, with the lan-
guage areas acting as a kind of central switchboard. There is also evidence of
close links between grammatical and lexical decits, which in turn suggests
that these two aspects of a speakers linguistic competence are closely inter-
twined.
Another important lesson to be learned from the research on aphasia is that
our capacity to use language is extremely resilient. In immature individuals,
language can survive the loss of the language areas or even of the entire left
hemisphere. In adults, such large-scale reorganisation is not possible, perhaps
because the regions which take over language processing in brain-damaged
children are already committed to other functions. However, there is evidence

that even adults are able to recruit new areas or make new connections to some
extent. Furthermore, adults are certainly able to compensate for the damage
suered by developing new language processing strategies. Both of these facts
lend further support to the claim that the architecture supporting the human
language faculty is very exible.
5 Language and other cognitive processes
1. The modularity hypothesis
One of the most controversial issues in contemporary psycholinguistics is the

extent to which our linguistic abilities depend on general purpose cognition.
In the 1980s and 1990s, many researchers (e.g. Ferreira and Clifton 1986;
Fodor 1983; Smith and Tsimpli 1995; Yamada 1990) proposed that language
is relatively independent from other cognitive capacities and is subserved by a
specialised cognitive module. This controversial hypothesis is based on two
types of arguments. First, it has been proposed that the specialised language
processor is faster, and therefore informationally encapsulated: the early stages
of language processing rely exclusively on lexical and syntactic knowledge,
and other kinds of knowledge can only be accessed after a preliminary syn-
tactic analysis has been completed. As we saw in Chapter 2, language process-
ing is indeed very fast; however, there is no strong evidence that the purely
linguistic processes i.e. lexical access and parsing outpace processes
which depend on non-linguistic real-world knowledge. A second, much more
powerful argument for the modularity hypothesis is the existence of apparent
dissociations between language (especially grammar) and other cognitive pro-
cesses: it seems that language can be seriously disrupted while other processes
remain relatively unaected, and, conversely, severe cognitive impairment may
be accompanied by little or no grammatical decit.
As observed in the introductory chapter, linguistics has developed in rela-
tive isolation from research on human psychology. The apparent dissociation
of language from other cognitive processes supports the status quo: it appears
to justify the separation of linguistic theory from other disciplines. We shall
see in this chapter, however, that the case for modularity has been overstated.
We begin with a brief summary of the evidence for the autonomy of language,
and then proceed to a discussion of some of the problems with the argument.
The textbook example of a relatively pure language impairment is Brocas
(non-uent) aphasia. Brocas aphasics have problems with just about every
aspect of language. They nd it dicult to access the words they need, and in
extreme cases, may be able to retrieve little more than concrete nouns. Their
Language and other cognitive processes 51
speech is slow, poorly articulated and agrammatic (they often leave out oblig-
atory function words and inections, at least in English and typologically
similar languages). They are also unable to use grammatical cues in compre-
hension: for example, they have problems with reversible passives such as The
boy is being chased by the girl, where it is impossible to tell from the meaning
of the words alone who does the chasing and who is being pursued. However,
the general cognitive abilities of Brocas aphasics are good: they are able to
solve complex problems (as long as these dont require the use of language),
they behave in socially appropriate ways, and they can make excellent use of
contextual information to infer the meaning of an utterance. In fact, their
pragmatic skills often allow them to compensate very successfully for their
grammatical decit, and for this reason Brocas aphasia was for a long time
regarded as a purely productive decit.
Decits which seem to aect language and language alone have also been
observed in acquisition. While the majority of developmental language disor-
ders can be attributed to hearing loss, general learning diculties or emotional
problems, some children develop language late and with considerable di-
culty in spite of the fact that they appear to be completely normal in every
other respect. Such children, diagnosed as suering from specic language
impairment, or SLI, usually have problems with several aspects of language
use, including articulation, vocabulary and grammar. Their utterances are
shorter and less uent than those of normal children; they frequently omit
function words and axes, and tend to avoid more complex grammatical
structures. They also make more errors of commission than other children,
and perform below age norms on comprehension tests. Most SLI children do
eventually learn to speak normally, although some residual problems often
remain; consequently, they are at risk for reading disorders, and tend to
perform less well than unimpaired children at school.
More surprisingly, language can apparently develop more or less normally
even in the face of severe cognitive impairment. Williamss syndrome (WS) is
a case in point. WS is a rare neurodevelopmental disorder characterised by a
complex of physical abnormalities (including heart and circulatory defects
and a characteristic pixie-like face) and mental retardation. Individuals with
WS typically have performance IQs of about 5060 and nd it dicult to nd
their way around or perform simple everyday tasks such as tying their shoe-
laces. Their linguistic skills, however, appear to be remarkably well preserved.
WS adolescents are uent and articulate quite loquacious, in fact. Their
utterances are generally well-formed and contain a variety of complex gram-
matical structures, including full passives, relative clauses and multiple embed-
ding. Here, for example, is how a 17-year-old WS girl with an IQ of 50
described the experience of having an MRI scan:
(1) There is a huge magnetic machine. It took a picture inside the brain.
You could talk but not move your head because that would ruin the
whole thing and they would have to start all over again. After its done
they show you your brain on a computer and they see how large it is.
And the machine on the other side of the room takes pictures from the
computer. They can take pictures instantly. Oh, it was very exciting!
(Bellugi et al. 1993: 183)
Individuals with WS also perform well on some formal tests of language

development. They tend to achieve high scores (sometimes above their mental
age) on vocabulary tests, both receptive and expressive. They also perform rel-
atively well on tasks tapping various grammatical abilities, including compre-
hension of passives, negatives, conditionals, and anaphoric and reexive
pronouns (Bellugi et al. 1994; Clahsen and Almazan 1998). Finally, they have
good metalinguistic skills: for example, they are able to dene familiar words
and to identify (and often correct) ungrammatical sentences (Bellugi et al.
1993, 1994). Thus, it is clear that their linguistic skills far outstrip their non-
verbal abilities.
We nd another striking example of the selective sparing of linguistic abil-
ities in Christopher, a linguistic savant studied by Smith and Tsimpli (1991,
1995). Although Christopher appears to have a normal verbal intelligence, his
non-verbal IQ scores are extremely low (from 40 to 76, depending on the test);
at age 30, he fails a theory-of-mind test which poses no problem for normal 4-
year-olds and conservation tasks which normal children solve at the age of 5;
and he must be institutionalised because he is unable to look after himself.
However, Christopher has an unusual talent for learning languages: in addi-
tion to his native English, he is able to read, write and communicate in some
fteen other languages.
Thus, language, and in particular grammar, can be signicantly impaired in
the absence of other cognitive problems, and, conversely, language can be dis-
rupted while other cognitive processes remain relatively unaected. Many
researchers argue that such a double dissociation shows that language is
mediated by dedicated neural circuits, and hence independent from other cog-
nitive processes and hardwired that is to say, genetically predetermined (see,
for example, Pinker 1999; Smith and Tsimpli 1995).
2. Problems with the double dissociation argument
Such far-reaching conclusions, however, are rather premature, for several

reasons. First, saying that some individuals have intelligence without lan-
guage while others have language without intelligence (cf. Pinker 1995) is a
vast oversimplication, as the observed dissociations between language and
other cognitive processes are partial. Global aphasia the total loss of the
ability to communicate by means of language is comparatively rare and is
apparently always accompanied by cognitive decits (van Mourik et al. 1992).
In most aphasic patients, some linguistic abilities are preserved (see Chapter
4), so it is inaccurate to say that they lack language; and while they usually
perform well on cognitive tests, their non-linguistic cognition is rarely, if ever,
completely unaected. A number of studies, for example, have reported that
aphasics obtained lower scores than other brain-damaged patients on non-
verbal intelligence tests (see Baldo et al. 2001; Hamsher 1998; Keil and
Kaszniak 2002; Lebrun and Hoops 1974), although it is not clear whether this
is a result of aphasia as such or of constructional apraxia (inability to copy or
construct complex gures), which often accompanies aphasia. Many aphasics
also have problems with relatively simple tasks which do not involve language,
such as using the right colours to ll in outline drawings, associating non-lin-
guistic sounds with pictures, showing what one does with various objects (e.g.
a toothbrush), and sorting. The co-occurrence of the linguistic problems with
such non-linguistic decits does not entail that there is a functional link it is
possible, for example, that these non-linguistic tasks are subserved by adjacent
areas of the brain which are damaged at the same time as the language areas
but it should make us careful in drawing conclusions about the presumed
independence of language from other cognitive processes.
The same point can be made even more forcefully with regard to language-
impaired children. Individuals with SLI have problems with language, but
they are certainly not languageless; and although, by denition, they have
normal or above-normal intelligence, there is ample evidence that various sub-
populations of children diagnosed with SLI suer from associated non-
linguistic decits. The most often reported impairments involve auditory
processing and short-term memory, but a variety of other problems have been
observed as well (Bishop 1992; Curtiss and Tallal 1991; Curtis et al. 1992;
Gathercole and Baddeley 1990; Hill 2001; Johnston 1997; Jones et al. 1994;
Merzenich et al. 1996; Montgomery 1995; Vargha-Khadem et al. 1995).
Again, the co-occurrence of language disorders with such non-linguistic de-
cits does not prove that the former are a consequence of the latter: it is pos-
sible that the children had been misdiagnosed, or that their non-linguistic
problems are a result of the language impairment (cf. Stromswold 1996).
However, the existence of such associated decits suggests that the term spe-
cic language impairment is a misnomer.
The case for language without intelligence is equally problematic. It is true
that the hyperlinguistic individuals mentioned earlier have low intelligence.
The WS subjects described in the literature usually have IQs of about 5060,
and Christophers full-scale IQ is probably between 70 and 80. However, this
does not mean that they are cognitive vegetables. As Tomasello (1995) points
out, IQ is a number derived by dividing a persons mental age by their chron-
ological age (and multiplying the result by 100). Thus, the IQ of 50 obtained
by the girl quoted in the preceding section means that, at 17, her performance
on an intelligence test was equivalent to that of an average 8-and-a-half-year-
old. The studies documenting the good performance of WS subjects on
linguistic tasks involved adolescents and young adults with mental ages
between 5 and 9. Since normal children in this age range speak uently and
understand passives, negatives, conditionals etc., we should not be surprised
by the fact that WS adolescents are also able to understand such structures.
The same argument also applies to Christopher, of course and in his case,
one might also add that his verbal IQ scores were within the normal range,
that he performed extremely well on a test of pragmatic inference, and that he
was able to teach himself foreign languages by reading grammar books. Thus,
describing people like Christopher or the WS patients studied by Bellugi and
her colleagues as lacking intelligence is simply preposterous.
Furthermore, it is important to note that none of these hyperlinguistic
individuals actually has normal language. People with WS show considerable
language delay in infancy (Bellugi et al. 1994; Paterson et al. 1999; Thal et al.
1989). They tend to make good progress in middle and late childhood, and by
adolescence, their linguistic skills are comparable to those of normally devel-
oping 67-year-old children (Karmilo-Smith et al. 1997; Volterra et al. 1996)
that is to say, about right for their mental age, although even adolescents
have diculties with complex syntactic structures (Grant et al. 2002). In other
words, the linguistic skills of individuals with WS are good only in compari-
son with their non-verbal abilities. This was demonstrated particularly dra-
matically by Stojanovik et al. (submitted), who, unlike other researchers,
directly compared WS and SLI individuals on the same tests. In line with pre-
vious research, they found dissociations. The SLI participants performed at
age-appropriate levels on standardised tests of non-linguistic cognition, and
below age levels on language tests. The WS participants, on the other hand,
performed well on language tests relative to their non-verbal cognition scores.
Crucially, however, they were no better than the SLI participants on the lan-
guage tasks (in fact, they were slightly worse on some measures). Given that
SLI is supposed to be a selective impairment of language, it makes no sense
to say that language skills in WS are intact.
Similar observations can be made about Christopher. We do not know
much about his early linguistic development (he was already 29 when Smith
and Tsimpli began their investigation), although we do know that he was a late
talker. His command of foreign languages, while impressive, is far from
perfect: in fact, he seems to have internalised very little syntactic knowledge,
although his vocabulary and morphology are very good. He tends to follow
English word order in spontaneous production and translation, and his gram-
maticality judgements are strongly inuenced by English. Smith and Tsimpli
conclude that Christophers syntax is basically English with a range of alter-
native veneers (1995: 122) and observe that it is possible that despite speak-
ing many languages, Christopher really has only one grammar (129).
On the other hand, argue Smith and Tsimpli, Christophers command of
his native English is virtually perfect. He performed at ceiling on TROG (Test
of Reception of Grammar), which shows grammatical knowledge equivalent
to at least that of a normal 11-year-old. Furthermore, he was able to make

very subtle distinctions, as well as correct ungrammatical sentences, on a
grammaticality judgement task. For example, he correctly identied the (a)
sentences below as ungrammatical, and oered the grammatical versions
given in (b):
(2) a. *This is the girl that I saw her with John at the cinema.
b. This is the girl that I saw with John at the cinema.
(3) a. *Mary believes the claim which John is a very intelligent man.
b. Mary believes the claim that John is a very intelligent man.
(4) a. *Which student do you think that could solve the problem?
b. Which student do you think could solve the problem?
Such subtle judgements clearly require sophisticated syntactic knowledge.

However, Christopher makes very little use of this knowledge in production.
His spontaneous conversation is generally monosyllabic. On the rare occa-
sions when he does produce longer stretches of discourse, his speech is mildly
repetitive and full of snatches that appear to have been memorised from text-
books (Smith and Tsimpli 1991: 327) and, one might add, not very native-
like (cf. example (5) below). Furthermore, he often produces ungrammatical
sentences when translating into English.
(5) IT How did you learn Hindi?

C From book.
IT Like Greek.
C Yes. Shall we try a bit of um exercises where, where the where you
want to go to a girl who doesnt speak Hindi and her mother says:
Are you are you permitted to go here because theres a lm on and
there and mother says a:j, a:j ek lm you a:p tu um kiya:
baza:r kya: zaba:r um now how do I write Hindi? Is it, is it this
way to write Hindi?
Smith and Tsimpli attribute these diculties to his non-linguistic cognitive

impairment. While this may account for the mistakes in translation (which is
without doubt a very demanding task), it does not seem like a plausible expla-
nation for his diculty with normal everyday conversation after all, WS ado-
lescents are uent conversationalists in spite of the fact that their cognitive
impairment is even more severe than Christophers.
In summary: it is true that in some people language is seriously impaired in
the absence of equally obvious problems with non-verbal cognition, while others
develop relatively good language in spite of severe cognitive decits. However,
such dissociations are partial. Hyperlinguistic individuals have unusually
good language, relative to their non-verbal abilities, while language-impaired
individuals have poor language skills, again relative to other cognitive domains.
In both groups, however, the decit is a matter of degree.
We now turn to a more important question: what exactly do these dissoci-
ations show? They do indicate that language is not merely an outgrowth of
general cognition. But then there is no such thing as general cognition: what
we have is a varied repertoire of cognitive skills and abilities. Some of these
(e.g. auditory short-term memory) are clearly implicated in language acquisi-
tion and processing, while others (e.g. the ability to nd ones way around in
an unfamiliar neighbourhood) are not. It is to be expected, then, that some
cognitive decits will have a profound eect on language performance, while
others will be of little or no consequence.
On the other hand, the dissociations do not show that grammar is an innate,
domain-specic and informationally encapsulated module indeed, the very
fact that they are partial entails that there are important links between verbal
and non-verbal cognition. Specically, the fact that both lexical and grammat-
ical development is signicantly delayed in virtually all mentally impaired chil-
dren including those who are described as hyperlinguistic suggests that
certain cognitive infrastructures, to borrow Elizabeth Batess (1994) term,
must be in place before grammatical development can begin. One should
stress, however, that these are fairly modest: equivalent, perhaps, to the mental
abilities of a normal 2-year-old (cf. Bates 1997). Furthermore, the fact that the
language of the mentally handicapped again including those who are hyper-
linguistic is also impaired, albeit sometimes to a lesser degree than other
aspects of their cognitive functioning, suggests that some grammatical abil-
ities require more elaborate cognitive infrastructures.
This brings us to the nal point: it is not accurate to say that language pro-
cessing in hyperlinguistic individuals is simply not quite as good as in
normals: there is evidence that it may also be qualitatively dierent. Several
researchers have noted that the spontaneous speech of people with WS often
includes strange word choices. For example, one of the young people studied
by Bellugi observed that The bees abort the beehive (that is to say, leave it);
another indicated that she would have to evacuate the glass (empty it) (Bellugi
et al. 1994: 32). Individuals with WS also tend to use more low-frequency
words than normals on word uency tasks, which prompted some researchers
to conclude that their mental lexicons are organised in an unusual (Bellugi et
al. 1994) or even deviant (Pinker 1991) fashion.
The language of children and adolescents with WS also shows some gram-
matical peculiarities. In contrast to normal children and adults, for example,
they perform better on high-frequency than on low-frequency regular verbs,
but show no frequency eects on irregulars (Thomas et al. 2001). Karmilo-
Smith and colleagues (1997) found that their French-speaking WS subjects
made signicantly more gender-agreement errors than controls, especially
when the experimental stimuli were nonce words rather than real words.
Particular diculties with agreement were also observed in the speech of an
Italian WS girl studied by Capirci et al. (1996). Capirci et al. also note that the
girl made syntactic errors that do not occur in the speech of normal children
at comparable stages of linguistic development. Finally, Karmilo-Smith et
al. (1998), using an online language task, found that their WS subjects showed
normal sensitivity to phrase structure violations and restrictions imposed by
an auxiliary on the form of the following verb, but not to lexical verb subcat-
egorisation frames.
Such ndings suggest that people with WS may acquire and process lan-
guage in a qualitatively dierent way (cf. Bellugi et al. 2000; Karmilo-Smith
1998; Karmilo-Smith et al. 1998). There is some support for this conclusion
from brain imaging studies, which have shown that WS individuals show dis-
tinctive ERP patterns which are not found in any other group (Bellugi et al.
1994; Neville et al. 1993). Furthermore, they show no left-hemisphere domi-
nance for language that is to say, unlike normal adults and older children,
they process language bilaterally.
As we saw earlier, Christophers language is also deviant, though apparently
in dierent ways. The dierence is most readily seen when we compare his
spontaneous speech with that of WS subjects. People with WS speak uently
and generally correctly and use a variety of complex syntactic structures.
Christophers conversation, in contrast, is monosyllabic and non-native-like.
On the other hand, Christopher performs much better than WS subjects on
formal tests of language development and on grammaticality judgement
tasks.
At the moment we know too little about the language abilities of hyperlin-
guistic individuals to draw any denite conclusions. It is possible, however,
that they learn and process language in qualitatively dierent ways from
normal people. If this is the case, their abilities oer further evidence of the
human brains ability to wire itself for language in dierent ways, according
to the mental resources available rather than of the existence of an innate,
informationally encapsulated language module.
6 Biological underpinnings
1. A genetically specied language module?
All normal humans, and no other species, have language. Many animals, of
course, have quite elaborate communication systems, but human language is
very dierent from animal languages they are not just simpler, but qualita-
tively dierent. Unlike other species, humans are able to refer to objects in
their environment (as well as to absent or imaginary entities) and combine
words and phrases to form novel messages. Furthermore, human languages,
even if they require some innate knowledge, are largely learned, since children
growing up in dierent linguistic communities end up speaking dierent lan-
guages. Last but not least, human languages, unlike the vocalisations of our
closest ape relatives the chimpanzees are under cortical control, and hence
conscious and deliberate. Chimpanzee vocalisations, in contrast, are bound:
they are triggered by a particular emotional state, are very dicult to suppress
in its presence, and cannot be produced in its absence (Goodall 1986).
Virtually all language researchers agree that language acquisition would be
impossible without some kind of innate structure: a atworm cannot learn
Flemish, or a chicken Chinese, however hard they and their teachers might try.
What is much less clear is what kind of innate structure is necessary for this
task, and the extent to which it is task-specic (devoted exclusively to language
learning). According to the generative tradition, which dominated linguistic
research for much of the latter half of the twentieth century, all humans are
endowed with Universal Grammar, a mental module specifying the univer-
sal properties of human languages and the parameters along which they can
vary, without which language acquisition would be impossible. Universal
Grammar is believed to have evolved after we separated from the apes and to
contain knowledge which is applicable exclusively to language learning in
other words, it is regarded as an adaptation which is both species- and task-
specic.
However, the idea of a language module or organ (cf. Anderson and
Lightfoot 2002; Chomsky 1975, 2000), though it may initially seem attractive,
becomes considerably less appealing when we consider the neuroanatomical
Biological underpinnings 59
evidence. One reason for this is that so far, despite intensive research, no one
has been able to nd this module. Anatomically, the human brain is not very
dierent from that of the great apes. It is considerably larger (about three times
the size of that of a chimpanzee), and it diers in the relative proportions in
that humans have more association cortex and a larger cerebellum. However,
there is no evidence that humans evolved new cortical areas or structures not
present in the apes (Elman et al. 1996; Mller 1992, 1996; Preuss 2000). The
so-called language areas (Brocas and Wernickes areas) have homologues in
ape brains (Preuss 2000) and are involved in other cognitive operations as well
(Fiez et al. 1995; Fox et al. 1988; Mazziotta et al. 1982).
One might object that the presence of homologous areas in ape brains does
not entail that the details of the microcircuitry are the same: it is possible that
in the course of human evolution, our brains have been rewired to encode
linguistic knowledge (and presumably knowledge underlying other uniquely
human abilities). However, there are several reasons for doubting that the
genome species the details of cortical connections, as opposed to the general
pattern.
First, there is an arithmetical problem which Mller (1996: 626) has dubbed
the problem of the poverty of the genetic code. The human brain contains
approximately 100 billion neurons. Assuming that each neuron is connected
to 1000 other neurons (a rather conservative estimate), we have 1014 (i.e.,
100,000,000,000,000) synapses. The human genome, however which encodes
information necessary for building the rest of the body as well contains only
about 24,000 genes (How many genes 2003), about 99 per cent of which are
shared with chimpanzees (Ebersberger et al. 2002). The disparity between
these two gures suggests that the details of cortical microcircuitry are prob-
ably not fully specied in the genome, but rather emerge through complex
interactions of genetic and environmental factors (cf. Elman et al. 1996).
Secondly, there is considerable evidence that cortical tissue is quite plastic
in that it can support a variety of dierent representations, depending on the
input it receives (Deacon 1997, 2000; Elman et al. 1996; Johnson 1998). We
have already seen in Chapter 4 that if the regions which are normally respon-
sible for linguistic processing are damaged early in infancy, language usually
develops in other parts of the brain. Similar reorganisations have also been
observed for other cognitive functions in many species. When a particular area
is deprived of its usual input, it may assume another function. For example, if
a monkeys nger is amputated, the area which would normally receive input
from it will be colonised by input from adjacent ngers; or parts of the audi-
tory cortex may take over visual processing if they do not receive input from
the ears. Such reassignment of function is most pronounced if the damage
takes place early in infancy, but also occurs to some degree in mature individ-
uals. Furthermore, when a living brain is surgically rewired so that a partic-
ular region receives input from a dierent source so that the auditory cortex
receives information from the eyes, for example the new recipient region
develops some characteristics of the normal recipient region (in our example,
the auditory cortex takes on visual representations). But perhaps the most dra-
matic evidence for plasticity has come from experiments involving transplan-
tation of fetal neural tissue to a dierent area of the brain in another animal
of the same species, or even a dierent species. Amazingly, the transplanted
neurons grow appropriate connections and become functionally integrated
with the host neurons (Deacon 1997, 2000), demonstrating that the genetic
program for growing the brain is not only exible, but also quite similar across
species. For obvious ethical reasons, such experiments have not been per-
formed on humans, but research on brain reorganisation in amputees and con-
genitally deaf or blind individuals is generally consistent with the animal data.
Moreover, there have been several apparently successful attempts to trans-
plant fetal pig neural cells into the brains of human patients suering from
Parkinsons disease in an eort to improve their motor control. A post-
mortem carried out after one such patient died revealed that the pig cells had
survived and had grown appropriate connections in the host brain (Deacon et
al. 1997).
It must be stressed that there are limits to this plasticity (cf. Johnson 1993,
1998). While it is possible for a lesioned brain to reorganise itself, particularly
if the lesion occurred early, this is by no means always the case: the results of
brain damage can be catastrophic. Similarly, while it is possible to transplant
fetal neural tissue to a dierent site in another brain, it is certainly not pos-
sible to transplant anything anywhere. Furthermore, we do not know how well
the transplanted tissue or rewired brains work. Although they do seem to take
on representations appropriate to the input they receive, they may not do so
as eciently as an intact brain would. Thus, the fact that brain tissue is fairly
plastic in the sense that it can take on a variety of representations does not rule
out a certain degree of specialisation: it is perfectly possible that some regions
are better suited for certain tasks than others. Even a fairly modest degree of
plasticity, however, is incompatible with the view that specic regions are pre-
destined to support a particular function because they encode the information
necessary for that function.
Thus, it seems that the genetic wiring diagrams for the brain are fairly
crude: fetal axons know the general area where they should terminate, but
not the specic cell. Patterns of connectivity which can support complex com-
putations such as language processing emerge later, largely as a result of
experience (cf. Elman et al. 1996; Johnson 1998). In what sense, then, can lan-
guage be said to be innate? To answer this question, it will be useful to appeal
to a taxonomy of ways to be innate oered by Elman et al. (1996; see also
Elman 1999). Elman et al. distinguish between three types of innateness:
1. Representational: the genome prespecies the neural microcircuitry

required for the activity, thus in eect endowing the organism with innate
knowledge.
2. Architectural: the genome species the general characteristics of the pro-

cessing system such as the types of neurons and their properties, the types
of neurotransmitters, the degree of interconnectivity in various regions of
the brain, and the overall patterns of connectivity (i.e. which brain regions
are connected and their characteristic sources of input and output).
3. Chronotopic: the genome controls the timing of the developmental pro-
cesses which build the brain, such as the number of cell divisions at partic-
ular stages of development and the patterns of synaptic growth and
pruning in various regions.
The evidence for cortical plasticity reviewed above argues against innately
specied cortical representations.1 This suggests that the innate biases that
construct a brain capable of acquiring language operate at the other two levels.
Architectural and chronotopic constraints require less genetically specied
information, so the small number of genes in the genome (relative to the
number of synapses in the brain) is not a problem. They operate indirectly, by
enabling certain processes and making certain outcomes highly probable, but
they also allow for variations after injury or when environmental conditions
are dierent. However, architectural and chronotopic constraints are not spe-
cic enough to encode details of linguistic knowledge: instead, they build a
brain that can perform certain types of computations and support certain
types of learning.
It follows that human adaptations to language involve innate information-
processing and learning biases rather than innate knowledge. We now turn to
the question of what these adaptations might be, and how they might have
evolved.
2. Human adaptations to language
2.1 Preadaptations
Our evolutionary history saw two periods of relatively rapid increase in brain
size. During the rst of these, which occurred between 2 million and 1.5
million years ago, the brains of our ancestors doubled in size from approxi-
mately 450 to 900 cubic centimetres, and during the second, which began
about half a million years ago and lasted until about 300,000 years ago, they
reached their present size of approximately 1350 cc. Both of these periods of
brain expansion appear to have coincided with changes in the upper respira-
tory tract which would have enabled our ancestors to produce a wider range
of sounds (Lock and Peters 1996), so it is likely that they correspond to major
developments in the evolution of language, perhaps the emergence of sym-
bolic communication and combinatorial syntax respectively.
However we interpret the fossil record, it is clear that language is a relatively
recent innovation. Since evolution is a highly conservative process which

works by modifying or combining existing systems rather than creating new
ones from scratch, the recent origin of language makes it virtually certain that
it builds on pre-existing abilities which evolved for other purposes. These
would have included, among others:
1. proto-phonological capacities, i.e. the ability to produce and recognise

sounds and to coordinate complex motor routines involving the articula-
tors;
2. proto-semantic capacities such as the ability to categorise percepts and to
construct complex concepts (as well as some elementary concepts and/or
feature detectors);
3. the capacity to form cross-modal associations (necessary to learn
formmeaning pairings);
4. the ability to store and retrieve such pairings when required.
All of these abilities are particularly well developed in humans, but they are
also found in apes (Lock and Colombo 1996; Premack 1980), so they can be
regarded as preadaptations for language. Since they are clearly useful for lan-
guage, it is reasonable to assume that, once humans acquired a rudimentary
linguistic system, they were selected for, and rened as a result: the vocal tract
became increasingly suitable for speech production, we became better at co-
ordinating motor routines, acquired more sophisticated conceptual abilities
and a better memory (especially for cross-modal associations), and so on.
However, before any of this could happen, our ancestors had to acquire an
ability without which it is impossible to have language as we know it: the
capacity for cultural learning.
2.2 Cultural learning and mind-reading

The capacity for cultural learning made possible the transmission of cultural
innovations, which in turn cleared the path for cumulative cultural evolution:
a process in which later generations do not have to reinvent everything from
scratch, but can build on the accomplishments of preceding generations.
Cumulative cultural evolution is a necessary condition for language:
without it, language would never get o the ground. One person, or even a
small group of people, cannot create a language: it takes a whole community
or, to be more exact, several cohorts of learner-inventors in a community.
We know this from research on the development of gestural communication
in the deaf. Deaf children deprived of sign language input develop a rudimen-
tary form of communication (homesign), but not a full language. The emer-
gence of a true language requires a critical mass of individuals who interact
on a regular basis (Kegl et al. 1999). It also requires time, as demonstrated by
Senghas and Coppola (2001), who documented how Nicaraguan Sign
Language emerged after a number of deaf children were brought together in

special schools and was grammaticalised by successive cohorts of learners
who built on the conventions established earlier.2
Cumulative cultural evolution (whether linguistic or non-linguistic) appears
to be a uniquely human trait, and relies on a characteristically human form of
learning: imitation (Tomasello 1999). Our ape relatives do develop certain
cultural practices such as eating termites in a particular way or using two
stones to crack nuts. They also can, and do, learn to use gestures communica-
tively. However, neither of these behaviours is acquired in the same way as
human cultural practices and gestural signs. To appreciate the dierence, it
will be useful to begin with a quick look at learning in chimpanzees and other
non-human primates.
Chimpanzee communicative gestures are learned through a process known
as ritualisation. They originate from a non-communicative behaviour
intended to accomplish a particular goal; in the course of ritualisation, the
behaviour gets abbreviated and comes to be used as a means of getting another
individual to do so something. For example, an infant chimpanzee might pull
the mothers arm to reach her nipple, and the mother will typically respond by
allowing him to suck. After the behaviour has been repeated a number of
times, the mother anticipates the infants actions and begins to respond as
soon as he has touched her arm. The baby in turn discovers that he doesnt
actually need to pull the mothers arm: all he needs to do is touch it. Thus, the
gesture becames abbreviated and acquires a meaning: Let me suck.
An important characteristic of ritualised gestures is that they are idiosyn-
cratic: they are used only by the two individuals who participated in the ritu-
alisation process. In fact, neither of the two individuals can be said to know
the gesture fully in the way humans know linguistic signs, because each knows
only half of it: in our example, the baby knows how to produce it, and the
mother how to interpret it, and they never reverse roles. Of course, in this case,
it would be inappropriate for the mother to ask to be allowed to suck; but such
one-sidedness is characteristic of gestures learned by ritualisation (Tomasello
2003).
Another form of social learning which is typical for chimpanzees is emula-
tion. In emulation, an animal observes another animal produce an eect, and
later learns to reproduce the eect. Emulation is thus similar to imitation, in
that both involve learning from a model. The crucial dierence is that in imi-
tation, the learner learns that a particular eect can be achieved by using a par-
ticular method; in emulation, the learner learns that a particular eect can be
achieved, but not the method the latter is learned later, typically by trial and
error. A good example of emulation is the acquisition of the nut-cracking
technique by chimpanzees living in the Tai forest, which involves placing a nut
on a hard stone and hitting it with another stone. The behaviour is peculiar to
the group living in Tai and is learned (other populations of chimps do not
crack nuts in this manner). Nuts are an important source of food for Tai
chimps, and hence juveniles have plenty of opportunity to observe adults

cracking nuts. In spite of this, they require years of practice before they master
the technique: young chimps spend a great deal of time hitting hammers
against anvils without putting a nut in between, striking nuts with a hammer
without rst placing them on an anvil, or striking nuts on an anvil with the
hand rather than a hammer (Boesch 1993; Boesch and Boesch-Acherman
2000). Thus, each learner basically rediscovers the invention for himself. Such
learning is slower, and not faithful enough to achieve cumulative cultural evo-
lution.
Learning by ritualisation and emulation also occurs in humans, but the
most characteristic form of social learning in our species is imitation. In imi-
tation, both the method and the result it produces are learned by observation.
Curiously, non-human primates are either unable to learn in this way, or their
abilities to do so are severely limited. Tomasello and colleagues demonstrate
this by means of a simple experiment (see Tomasello 1999: 32). They removed
one chimpanzee from the group and taught her to produce a particular
gesture, for which she was rewarded with food. They then returned the animal
to the group, where other individuals could observe her perform the gesture
and get her reward; but although they clearly wanted the food, none imitated
the gesture. Other researchers (e.g. Whiten et al. 1996; Celli et al. 2001) have
argued that some non-human primates, notably chimpanzees, are capable of
some learning by imitation. Crucially, however, none of the studies that attrib-
ute imitative abilities to chimpanzees involved imitation of communicative
signs.
Tomasello (1999) concludes that there are two critical dierences between
social learning in humans and other primates. First, humans appear to be much
more tuned to the details of the models behaviour, and aim to reproduce them
faithfully (even irrelevant details which are not necessary for the successful exe-
cution of the action: Nagell et al. 1993). Secondly, while other primates seem
to focus on the environmental eects of the actions they observe, human learn-
ers focus on the models goals. This is because we perceive other humans as
intentional agents, and naturally attribute intentions to them: on seeing
someone struggling with a lid on a box, a human observer cannot help but per-
ceive that person as a trying to open the box. Human infants as young as 18
months imitate actions that the model tried to perform but failed (in other
words, they can work out what the models goals were even if they had not seen
the action succeed: Meltzo 1995), but they tend not to imitate unintentional
actions (Carpener et al. 1998). This, Tomasello argues, is because humans have
a theory of mind: we know that other humans are intentional agents like our-
selves, and we can put ourselves in their shoes. Because of this, we perceive
other people as having intentions (including communicative intentions) and
are very good at working out what those intentions might be.
These two characteristics of human cultural learning are highly relevant to
the development of language, both in the individual and in the wider histori-
cal sense. The propensity for faithful imitation ensures that all individuals in a
community acquire (more or less) the same signs. Our mind-reading skills, on
the other hand, ensure that we attach roughly the same meanings to these
signs.3 Both of these abilities are a necessary prerequisite for language.
2.3 Cortical control over vocalisations and coordinating action sequences

The vocalisations of chimpanzees and other non-human primates are innate
responses triggered by a particular emotional state, rather like laughter or
sobbing in humans. They are controlled by the viscero-motor system, an auto-
matic response system connected with arousal states. Human speech, in con-
trast, is under voluntary control: we can decide what to say in a particular
situation or whether to say anything at all (although the latter ability may be
severely impaired in some individuals). This means that in the course of
hominid evolution, control over the oral musculature has shifted to the
cortex.4
Cortical control, apart from enabling us to say what we want when we want,
had two further consequences. First, it gave us the ability to learn new vocal-
isations, thus vastly increasing our vocal repertoire and, concomitantly, the
number of distinct messages that could be conveyed. Secondly, it made it pos-
sible to harness other cortical systems notably those for coordinating
complex motor routines for use in communication, thus enabling our ances-
tors to combine simple vocalisations to form more complex ones and estab-
lishing the foundations for syntax.
This suggests that syntax (or at least some aspects of syntax) might be a spe-
cialisation of the more general ability to coordinate complex motor routines
and combine units hierarchically (cf. Lieberman 1991, 1998, 2002). Several
lines of evidence seem to support this conclusion. First, imaging studies show
that the same regions of the brain notably, Brocas area play a vital role in
syntax, speech motor control, and programming sequences of manual move-
ments (Fox et al. 1988; Kimura 1993). This is corroborated by research on de-
cits resulting from brain injury. As noted in Chapter 4, Brocas aphasia is
frequently (though not always) accompanied by constructional apraxia, which
suggests that the regions responsible for coordinating language use and pur-
posive manual movements partly overlap.5 Moreover, Grossman (1980) noted
that even in the absence of apraxia, Brocas aphasics have a selective decit in
reconstructing the hierarchical organisation of an abstract diagram from
memory. Motor decits are also very common in children suering from SLI
(Bishop 2002; Hill 2001). Conversely, syntactic decits have been observed in
patients suering from motor diseases such as Parkinsons (Lieberman et al.
1990, 1992). While such associations are not perfect (not all syntactic decits
are accompanied by a motor decit, and not all people with impaired motor
control also have language problems), their coincidence is high enough to
suggest a common underlying problem.
Additional evidence for a possible link between motor planning and lan-
guage comes from comparative studies of humans and other primates. All pri-
mates are able to learn sequences of actions or stimuli, which suggests that it
is an ability that we inherited from our common ancestors. However, a more
detailed analysis of the performance of humans and modern-day non-human
primates reveals that there are some subtle dierences in how they approach
the task: specically, humans plan the whole complex sequence before execut-
ing it, whereas non-human primates appear to program each stage indepen-
dently (although chimpanzees show some ability to plan ahead; see Conway
and Christiansen 2001 for a discussion).
The ability to program and execute a whole sequence of actions as one unit
may have developed in humans as an adaptation to throwing (Calvin 1983,
1993). Accurate throwing is useful for chasing away predators and for hunting.
However, it is quite a demanding task, in that it requires the coordination of
the action of many muscles and very precise timing. Furthermore, because it
is a very brief action, there isnt enough time for the brain to respond to feed-
back from the moving arm; therefore, unlike slower actions such as climbing
a tree to reach a piece of fruit, the whole motor sequence for throwing must
be planned before the launching phase is initiated. As Calvin put it:
your arm is an unguided missile shortly after the throw has begun. You
must plan perfectly as you get set to throw, create a chain of muscle
commands, all ready to be executed in exactly the right order . . . You
need something like a serial buer memory in which to load up all the
muscle commands in the right order and with the right timing relative to
one another. (1993: 234)
Calvin also notes that because neurons are rather inaccurate time-keepers, in
order to achieve the precise timing necessary to hit the target at any distance,
the planning mechanism must average the output of a very large number of
neurons, and hence it is possible that the advantages of accurate throwing
might have driven the selection for a bigger brain in our ancestors.
Whatever its evolutionary origins, the ability to package a series of actions
as a single unit allows more elaborate planning, since the complex unit itself
can be incorporated into a larger action plan. It could also be exapted for lin-
guistic use: a sequence of words functioning as a single unit that is to say, a
phrase can be incorporated into a larger unit, thus giving rise to hierarchi-
cal syntax. It is interesting to note, then, that while the giant apes are able to
learn hierarchically structured action sequences, their ability to do so is quite
limited similar to that of a 2-year-old human child (Conway and
Christiansen 2001).6 This appears to parallel the upper limit of their linguis-
tic abilities. Attempts to teach language-like systems to captive apes have
shown that they can learn simple ordering rules, but they do not seem to be
able to combine phrases hierarchically.
3. Language adaptations to humans
While it is undeniable that in the course of human evolution our brains and
vocal apparatus adapted to the exigencies of linguistic communication, we
should not lose sight of the fact that languages, too, are products of evolution.
The emergence of language is a process in which individual speakers make
innovations which either do or do not spread through the community. Whether
or not a particular variant will spread depends at least in part on its commu-
nicative tness. Other things being equal, speakers naturally prefer structures
which are easier to produce; and through failure to communicate, they learn to
avoid structures which are dicult to understand.7 This means that when there
is more than one structure expressing a particular function, the structure which
is more dicult to process will be used less frequently. The next generation of
speakers will have the same processing biases, but also fewer opportunities to
learn the more dicult structure. This will result in a further drop in frequency
of the more dicult variant, until it eventually disappears.
In other words, the story of the emergence of language is one of co-evolu-
tion: humans adapted to language, but language (or rather languages) also
adapted to humans. Indeed, there are reasons to think that languages did most
of the evolving. As Deacon (1997) points out, languages need humans more
than humans need languages (if it came to the crunch, we could probably
survive without them; but they need our brains to survive and reproduce) so
the selectional pressures on languages were stronger. Furthermore, cultural
change is several orders of magnitude faster than biological change. Any sub-
stantial restructuring of the brain requires hundreds of thousands or even mil-
lions of years. Grammars, on the other hand, can change radically in a few
centuries, and lexical innovations spread in a matter of years and, in this era
of mass media, sometimes in a matter of months or even weeks.
One particularly good example of language adjusting to humans is the phe-
nomenon of categorical perception of speech sounds. When presented with a
range of stimuli which dier along some continuous parameter, such as voice-
onset time (VOT), people tend to perceive sounds up to a certain point on the
continuum as belonging to one category and sounds that fall beyond that
point as belonging to another category. For example, stops with a VOT of less
than 35 ms are perceived as voiced, while those with a greater VOT are per-
ceived as voiceless. Thus, an articially produced sound that is intermediate
between a prototypical /t/ and a prototypical /d/ is not perceived as something
in between the two phonemes, but as either one or the other, depending on
whether the VOT is less than or more than 35 ms. We know that at least some
such perceptual biases are innate because they are present in very young
infants, even when the ambient language does not make use of a particular
phonemic contrast.
Our tendency to perceive speech sounds categorically is useful from a
communicative point of view, since it allows ecient coding of phonemic
distinctions, and used to be regarded as a prime example of an innate special-

isation for language or, to be more precise, for speech. However, categorical
perception is not for language, or speech. Some non-speech sounds are also
perceived categorically (Premack 1980). More importantly, categorical percep-
tion of human speech sounds is found in a number of other species, including
chinchillas and Japanese quail (Kluender et al. 1987; Kuhl 1981; Miller and
Jusczyk 1989; Springer 1979). In other words, the existence of categorical per-
ception of speech sounds in humans is not evidence that our species adapted
to the exigencies of vocal communication, but rather that language adapted to
our pre-existing auditory biases.8
Thinking in terms of languages adapting to humans has profound implica-
tions for the way we think of language universals: it suggests that it might be
more revealing to view them not as a set of constraints dening a possible
human language encoded in the genome, but rather as the result of convergent
evolution. Deacon (1997) develops this idea at length, concluding that:
universal rules and implicit axioms of grammar arent really stored or

located anywhere, and in an important sense, they are not determined at
all. Instead, I want to suggest the radical possibility that they have
emerged spontaneously and independently in each evolving language, in
response to universal biases in the selection processes aecting language
transmission. They are convergent features of language evolution in the
same way that the dorsal ns of sharks, ichthyosaurs, and dolphins are
independent convergent adaptations of aquatic species. Like their bio-
logical counterparts, these structural commonalities present in all lan-
guages have each arisen in response to constraints imposed by a common
adaptive context. Some of the sources of universal selection on the evo-
lution of language structures include immature learning biases, the con-
straints of human vocal articulation and hearing, and the requirements
of symbolic reference . . . Because of these incessant inuences, languages
independently come to resemble one another, not in detail, but in terms
of certain general structural properties, and any disruption that under-
mines a languages t with its host will be selected against, leading to
reconvergence on universal patterns. (Deacon 1997: 116)
There is considerable evidence for such external forces acting on and

shaping grammars. Functional linguists have long argued that many cross-
linguistic patterns arise as a result of discourse pressures (Du Bois 1987; Foley
and Van Valin 1984) and a general preference for iconic structures (Givn
1989; Haiman 1985). Even more direct are pressures on the processing system,
which appear to be responsible for constraints on long-distance dependency
constructions (Kluender 1998; Kluender and Kutas 1993) and many word
order universals (Hawkins 1994; Kirby 1998). One should also note in this
connection that absolute universals properties which are shared by all lan-
guages are hard to nd, and they tend to be rather uninteresting: they are
statements like all languages have vowels and consonants, all languages have
nouns (words prototypically referring to things) and relational words, all lan-
guages have recursion. The more meaty universals are either statistical (most
languages have property X) or implicational (if a language has property X,
then it also has property Y). This seems more compatible with the view that
universals are a result of languages converging on certain solutions than with
the view that universals are innate constraints dening a possible human lan-
guage.
4. Universal Grammar again
4.1 Problems with the poverty-of-the-stimulus argument

When discussing human adaptations to language earlier in this chapter, I
refrained from mentioning any innate syntactic knowledge, or even any spe-
cically syntactic specialisations. This is not to suggest that such adaptations
did not take place, but simply that we do not know enough about the evolu-
tionary origin or the genetic basis of syntax to constrain linguistic theory in
any meaningful way. The argument for an innate Universal Grammar is essen-
tially a linguistic one. However, because it has played such an important role
in contemporary linguistic theory, it is tting to conclude this chapter with
some observations about its validity and explanatory power.
The existence of Universal Grammar is inferred from two premises:
1. Human languages have certain (universal) properties (e.g. restrictions on

movement such as the Subjacency Condition).
2. These properties cannot be learned because the relevant information is not
available in the input.
Therefore, some linguists have argued, knowledge of these properties must be

innate. This, in essence, is the poverty-of-the-stimulus argument, the most
powerful and direct argument for the existence of innate syntactic knowledge
(Chomsky 1986; Crain 1991; Lightfoot 1989). Other sources of evidence
which have been marshalled in favour of the innateness hypothesis (e.g. the
speed and alleged uniformity of language acquisition, dissociations between
language and other cognitive abilities, the fact that some language impair-
ments are inherited) play only a subsidiary role: while they do suggest the exis-
tence of biological adaptations to language, they tell us very little about the
nature of these adaptations specically, they do not necessarily imply the
existence of innate linguistic principles (as opposed to, for example, innate
learning biases or an enhanced ability to process hierarchically organised
structures).
To appreciate the validity of the poverty-of-the-stimulus argument, it is

important to note that it relies on certain assumptions about the properties of
the adult system, the kind of information that is available to learners, and the
abilities that they bring to the learning task. It is to these that we now turn.
4.1.1 Properties of the adult system

The argument from the poverty of the stimulus is predicated upon the claim
that all speakers have the abstract linguistic knowledge embodied in the prin-
ciples of Universal Grammar. However, this supposed universality of abstract
linguistic knowledge is an assumption, not an empirically established result.
It is true that virtually all neurologically unimpaired adults are procient lan-
guage users. This, however, does not entail that they are all equally procient,
or that they all rely on the same abstract rules; and there is accumulating evi-
dence that individuals with little formal schooling have considerable diculty
processing complex syntactic structures (Chipere 2003; Dabrowska 1997; see
also the evidence for shallow processing in Chapter 2).
Another important point to bear in mind is that many of the principles
which are supposed to constitute our innate language endowment are theory
internal. As Tomasello put it:
Many of the Generative Grammar structures that are found in English

can be found in other languages if it is generative grammarians who
are doing the looking. But these structures may not be found by linguists
of other theoretical persuasions because the structures are dened
dierently, or not recognized at all, in other linguistic theories.
(Tomasello 1995: 138)
The fact that restrictions on movement cannot be learned (if indeed this is a
fact) does not entail that they must therefore be innate: it could also mean that
we need a better linguistic theory one which does not require movement. Of
course, linguistic theory still needs to account for the dependency phenomena
that movement rules were postulated to explain; the point here is that the
alleged unlearnability of a particular grammatical rule or principle cannot be
taken as evidence for the innateness of Universal Grammar unless we have
independent evidence for the psychological reality of that rule or principle.
4.1.2 Information available in input

The second premise of the argument from the poverty of the stimulus is that
certain crucial kinds of information are not available to the learner. The
problem is that this is often asserted rather than actually demonstrated.
Consider Chomskys favourite illustration, the fact that children seem to
acquire a relatively complex structure-dependent rule of auxiliary fronting
(move the rst auxiliary after the subject to the beginning of the sentence)
rather than a simpler structure-independent rule (move the rst auxiliary to
the beginning of the sentence): thus, when forming interrogatives correspond-

ing to example (1), English speakers produce (3) rather than (2), which is what
the structure-independent rule would produce.
(1) Everything that has been damaged will be paid for.

(2) *Has everything that been damaged will be paid for?
(3) Will everything that has been damaged be paid for?
This, Chomsky asserts, occurs in spite of the fact that utterances that would
enable them to distinguish between the two hypotheses are so rare that a
person might go through much or all of his life without ever having been
exposed to the relevant evidence (Chomsky 1980a: 40). However, the asser-
tion is not supported by any data on how frequent or infrequent the relevant
utterances actually are, either here or in the many other subsequent rehash-
ings of the argument. Other linguists have taken the trouble to look at the
corpus data, and found that relevant examples constitute about 1 per cent of
all interrogative utterances in a wide range of corpora (Pullum and Scholz
2002). Now 1 per cent may not seem like a lot but since children are exposed
to vast amounts of input (cf. Chapters 2 and 3), it is 1 per cent of a very large
number: Pullum and Scholz estimate that by age 3, a typical child will have
heard between 7,500 and 22,500 relevant instances.9
The tendency to assert that some aspect of linguistic knowledge could not
possibly be learned from the input, without providing any supporting evi-
dence, is endemic in the nativist literature (see Pullum and Scholz 2002 for
further examples). This seriously undermines the poverty-of-the-stimulus
argument, since there is accumulating evidence that the information available
in the input is richer than previously thought (Redington and Chater 1998).
For example, lexical co-occurrence patterns can be used to classify words into
grammatical categories, and prosodic phrasing provides clues about constit-
uency. Also, in spite of frequent claims to the contrary, it appears that chil-
dren receive plenty of negative evidence, that is to say, evidence that an
utterance they have just produced is ill-formed by adult standards (see
Chouinard and Clark 2003 for an up-to-date discussion). It is true that some
kinds of knowledge (e.g. properties of the various kinds of empty categories)
cannot be acquired from the input available to children. But again, this does
not necessarily mean that the child must have innate knowledge of these: an
alternative conclusion would be that we need to look for ways of accounting
for the same phenomena which do not postulate such invisible and inaudible
constructs.
4.1.3 The childs language learning abilities

However rich innate linguistic knowledge may be, there are many idiosyncratic
properties of individual languages which it cannot specify: the actual words
in the language and their meanings, the grammatical markers and their
functions, subcategorisation requirements of particular verbs, various quirky

constructions which violate the usual grammatical patterns, morphological
classes, and so on. All of these must be acquired using some mechanism other
than setting the parameters of Universal Grammar.
Many of these language-specic aspects of linguistic knowledge are
abstract and quite complex in fact, they appear to be every bit as complex as
the core properties which are deemed to be specied in Universal Grammar
(Culicover 1999; Dabrowska 2000; see also next chapter for a discussion of the
complexities of lexical meaning and Chapter 8 on knowledge about the distri-
bution of inectional endings). It follows, therefore, that the learning mecha-
nisms necessary to acquire language-specic knowledge must be quite
powerful possibly powerful enough to learn at least some aspects of core
grammar.
4.2 Some new questions

Arguments for an innate Universal Grammar often take the form of a chal-
lenge to anti-nativist researchers to explain some particularly esoteric aspect
of grammatical knowledge. For example, Bickerton (1996) proposes that stu-
dents of human behaviour who believe in the creation or invention of lan-
guage should be required to take an examination in which they would have to
answer questions such as the following:
Question 1: Explain how, and why, the inventors of language arranged

things so that John wants someone to work for means John wants
someone such that he, John, can work for that person while John wants
someone to work for him means John wants someone such that that
person will work for him, John. State how you yourself learned the
reversal of meaning in the subordinate clause and show how its inven-
tion was culturally and/or biologically adaptive. (Bickerton 1996: 34)
Bickertons formulation is a caricature of the invention view. Those who

maintain that language was invented do not claim that it was designed at the
drawing board, but that it emerged gradually, over a number of generations,
as a consequence of being used to solve specic communicative problems.
However, Bickerton does make a valid point: namely, that existing empiri-
cist/constructivist theories of language acquisition have not dealt with many
issues which are central to the generative enterprise, such as how people use
grammatical cues to recover pronoun reference and missing arguments. The
problem is that the generativists also fail to provide a satisfactory explanation
of how the relevant knowledge is acquired. This is easiest to see if we turn the
tables on Bickerton and ask him to take the nativist version of the same exam.
The rst question would be the following:
Question 1: Explain how evolution arranged things so that John wants

someone to work for means John wants someone such that he, John, can
work for that person while John wants someone to work for him means
John wants someone such that that person will work for him, John.
State how, and where, the reversal of meaning in the subordinate clause
is encoded in the human genome and how and by what physical and/or
chemical processes it comes to be represented in the human brain.
Posing the question in this way is just as unfair to the nativists as Bickertons
original question was to the opposite camp. No sane nativist believes that the
dierence in meaning exemplied in Bickertons sentences is directly encoded
in our genes. But this raises the question of what actually is encoded: does the
human genome contain statements to the eect that a trace must be properly
governed? Does it contain instructions for nding traces? What form do these
take? The point, of course, is that proclaiming that some aspect of our linguis-
tic knowledge is innate does not answer the question of how we come to have
it it merely raises three new ones:
1. How did our species come to have this innate knowledge?

2. How is the knowledge encoded in our genes, and how does the genetic
information translate into linguistic representations in the brain?
3. How do learners access innate knowledge when constructing a grammar of
the language they hear?
As we will see shortly, all of these questions have yet to receive a satisfactory
answer.
4.2.1 Evolution
For someone who attaches so much importance to innate linguistic knowl-
edge, Chomsky is remarkably reticent about how the capacity for language
evolved. In 1980, he side-steps the issue by stating that:
These skills [i.e., the ability to acquire language] may well have arisen as
a concomitant of structural properties of the brain that developed for
other reasons. Suppose that there was selection for bigger brains, more
cortical surface, hemispheric specialisation for analytic processing, or
many other structural properties that can be imagined. The brain that
evolved might well have all sorts of special properties that are not indi-
vidually selected; there would be no miracle in this, but only the normal
workings of evolution. We have no idea, at present, how physical laws
apply when 1010 neurons are placed in an object the size of a basketball,
under the special conditions that arose during human evolution.
(Chomsky 1980b: 321)
More recently, he has expressed a similar opinion:
Perhaps these [properties of language] are simply emergent physical

properties of a brain that reaches a certain level of complexity under the
specic conditions of human evolution. (Chomsky 1991: 50)
Other researchers (e.g. Calvin and Bickerton 2000; Jackendo 2002; Pinker
and Bloom 1990) have been considerably more explicit. However, although the
scenarios they have proposed are not unreasonable, they are still very sketchy,
mutually incompatible, and, worst of all, unveriable. This is unavoidable:
since languages do not fossilise, accounts of language evolution are necessar-
ily little more than scientic just-so stories.
4.2.2 Genetics and neurogenesis

We also know very little about which genes are relevant to the development of
language, let alone about how they actually do their job, although in this case
we are almost certain to learn a great deal more in the future. The most prom-
ising avenue of research is the genetics of developmental language disorders.
We already know from twin studies that such disorders do indeed involve a sig-
nicant genetic component, since concordance rates for linguistic problems
are considerably higher for identical twins (who share all their genes) than for
fraternal twins (who share only 50 per cent of their genes) that is to say, given
one twin with a language impairment, the likelihood that the other twin will
also be impaired is much higher if they are identical than if they are fraternal
(Bishop et al. 1995; Stromswold 2001; Tomblin and Buckwalter 1998).
However, there is little research linking specic linguistic decits to specic
genes or groups of genes. The rst and, at the time of writing, only unam-
biguous link of this kind was established by Lai et al. (2001), who showed that
the severe speech and language diculties found in many members of the so-
called KE family were associated with a mutation in one particular gene,
FOXP2.10 This discovery is sometimes regarded as providing direct support for
the claim that Universal Grammar is encoded in the human genome. There are,
however, several reasons to be sceptical about such an interpretation. First, the
eects of the genetic anomaly in the KE family are relatively non-specic: the
aected members have problems with pronunciation, morphology and syntax
as well as with sequencing ne non-linguistic movements of the mouth and face
muscles; and their non-verbal IQs are, on average, 18 points lower than those
of non-aected members, though still within the normal range (Vargha-
Khadem et al. 1995). Secondly, FOXP2 is also found in other mammals, and
the human version of the gene diers at only three amino-acid positions from
the corresponding mouse gene, and at only two positions from the analogous
gene in chimpanzees, gorillas and rhesus macaques (Enard et al. 2002). Thirdly,
the gene appears to be involved in the development of the lungs, heart and gut
as well as several regions of the brain (Marcus and Fisher 2003).
Thus, while something about the normal version of human FOXP2 seems to
help an individual to develop language, the gene itself does not contain a blue-
print for grammar. FOXP2 is a transcription factor, or a gene which regulates
the functioning of other genes. It is possible that these other genes are respon-
sible for encoding the principles of Universal Grammar into the developing
brain but if so, they have not been found yet. Another possibility is that FOXP2
exerts its inuence in a much more roundabout way, for example by controlling
the amount of brain tissue available for sequencing ne facial movements.
4.2.3 Accessing innate knowledge

To be able to use innate knowledge about syntactic categories and relations,
children must be able to identify instances of these categories in the language
they hear. However, they cannot rely on some invariant features of the input
to do this, since the relevant categories are highly abstract and do not share
any perceptual features.11 This raises the question of how children link the
speech they hear with innate syntactic knowledge a fundamental problem
for nativist theories of acquisition which has received relatively little attention
in the literature, and, as yet, no satisfactory solution.
The most serious attempt to deal with the linking problem is Pinkers
semantic bootstrapping hypothesis (Pinker 1984, 1987). Pinker proposed that,
in addition to innate syntactic categories such as noun, subject and sister
of X, children are also equipped with innate semantic categories such as
thing, agent and argument of X, as well as innate linking rules connecting
the two (words for things are nouns, agents are normally subjects and so
on). Linking rules make it possible for children to identify examples of the rel-
evant syntactic categories and relations. Once they have done this, children can
induce their language-specic distributional characteristics, which in turn
enables them to identify other category members, such as abstract nouns,
passive subjects and so on.
There are several problems with Pinkers proposal. First, it requires further
innate knowledge in addition to Universal Grammar. Secondly, because it
exploits correlations between semantic and syntactic categories, it requires
such correlations to be universal but, as Pinker himself acknowledges, they
are not (for example, agents do not match up with prototypical subjects in
ergative languages, so in these languages innate linking rules would not only
not help, but actually be a hindrance). Thirdly, as Tomasello (2000) points out,
the theorys empirical predictions are incorrect: childrens earliest subjects
include some agents, but also many other semantic roles.
5. Conclusion
There is a growing consensus that language an evolutionary late-comer

piggybacks on structures which originally evolved for other purposes (Hauser
et al. 2002; Lieberman 1991, 1998; Tomasello 1999; Worden 1998), such as
producing and recognising sounds, planning motor sequences, categorisation,
storage and retrieval of information, predicting the actions of other individ-
uals, and cultural learning. Many researchers maintain that we also have some
language-specic adaptations, although there is considerable disagreement
about what these are and how they may have evolved. However, it should be
clear that, for reasons discussed earlier, a good theory of language will keep
such stipulations to a bare minimum. Furthermore, they should be treated as
tentative hypotheses to be conrmed or rejected by biological research rather
than as central tenets of the theory.
Notes
1. Humans do have some innate representations. For example, human infants

preferentially look at face-like stimuli, suggesting that they have an innate
representation of a human face. The latter is, however, highly schematic
(something like an oval with three blobs corresponding to the eyes and the
mouth), and appears to be encoded in subcortical structures, not the corti-
cal structures that are involved in face processing in adults (Johnson 1998).
Its function, it seems, is to ensure that infants pay attention to human faces,
which guarantees that the cortical system responsible for face recognition
in mature individuals has plenty of opportunity to learn about them.
2. Similar conclusions can be drawn from research on the formation and
subsequent development of creole languages (see e.g. Mhlhusler 1986).
3. The ubiquity of imitation does not, of course, rule out the possibility of
innovation. If the linguistic system is to develop, at least some members
of the community must occasionally produce novel signs or novel sign
combinations. Note, however, that for the innovation to become part of
the shared system, it must be imitated by others.
4. The shift is not complete, so oral muscles in humans are actually under
dual control. The viscero-motor system continues to control the more
primitive vocalisations such as laughter or crying; it is for this reason that
laughter and crying are dicult to suppress in certain emotional states
and even more dicult to fake convincingly.
5. It is also possible that the two regions are simply located next to each
other, and consequently both are often damaged as a result of the same
injury. However, there is evidence that the same regions are involved in
both functions. Electrical stimulation studies conducted by Ojemann and
his team, for example, found that there was 86 per cent overlap in the sites
where the sequencing of oral-facial movements and phoneme perception
was disrupted (Calvin and Ojemann 1994).
6. They are also not very good at throwing, which supports Calvins sce-
nario.
7. Humans are very good at monitoring the interlocutors understanding

and adjusting their speech accordingly. For example, adults, and even chil-
dren as young as 4 or 5, produce simplied, redundant speech when con-
versing with children under 3 (Snow 1986).
8. Some phonemic distinctions, for example the contrast between voiced and
prevoiced sounds found in Thai, are acquired. However, languages across
the world prefer those distinctions to which we are innately sensitive, so
the voiced/voiceless contrast is used phonemically much more frequently
than the voiced/prevoiced contrast.
9. Note, too, that the whole argument rests on the assumption that interrog-
atives are formed by moving the auxiliary from the position it occupies in
declaratives. Learners would also avoid making the error in (2) if they
learned a question-formation rule which did not involve movement, for
example a simple template like AUX + Subject NP + nonnite VP. As
shown in Chapter 9, there is considerable evidence that children do in fact
acquire such templates.
10. Recent genome-wide scans found four more regions on four dierent
chromosomes that appear to be involved in another kind of language dis-
order (Fisher et al. 2003). However, the specic genes responsible for the
disorder have not yet been identied.
11. Note that the evolution of Universal Grammar raises an analogous
problem. All known examples of innate and domain-specic knowledge
have some physical invariant (a motor program or a perceptual represen-
tation) which gives evolution something concrete to work on. However,
because the formmeaning mapping in human languages is arbitrary, it
diers from language to language, and it is not clear what could be
selected for in the course of evolution.
Part II
The Building Blocks of Language
7 Words
The standard argument for the innateness of syntax goes something like this:
children acquire the grammar of their native language remarkably quickly, on
minimal exposure, and with no overt teaching. Moreover, they go through the
same developmental stages and make relatively few errors. Most importantly,
some aspects of language (e.g. knowledge about empty categories, constraints
on movement) are acquired without any experience. Since no known learning
theory could begin to explain how this happens, linguistic knowledge must be
innate.
With this argument in mind, consider the following facts about lexical devel-
opment:
1. In early and middle childhood, children learn about ten new words a day
approximately one in every waking hour (Carey 1978; Miller 1986; Miller
and Gildea 1987).
2. Words are often learned after a few presentations in ambiguous conditions,
and sometimes after only one presentation (Carey 1978; Dickinson 1984;
Heibeck and Markman 1987; Markson and Bloom 1997).
3. Most words are learned without overt teaching, and attempts to teach
vocabulary at school sometimes lead to very strange results.1
4. Lexical development exhibits some striking regularities. For example,
colour terms are acquired at a specic maturational stage; and dimensional
adjectives and other spatial terms are acquired in a fairly uniform order.
5. Most words are acquired without eort and almost without errors, in spite
of the fact that lexical meanings are often very dicult to pin down. (Many
dictionary denitions are grossly inadequate, and even philosophers nd it
dicult to explicate our tacit knowledge of everyday concepts.)
6. Children can acquire words from domains of which they have no experi-
ence. For example, blind children have been shown to acquire visual terms
without much diculty (Landau and Gleitman 1985).
Applying the same logic to these facts, the conclusion seems inescapable:
lexical knowledge, or at least a signicant part of it, must be innate. In fact,
this is precisely the conclusion drawn by Chomsky:
Barring miracles, this means that the concepts must be essentially avail-
able prior to experience, in something like their full intricacy. Children
must be basically acquiring labels for concepts they already have.
(Chomsky 1991: 29)
Piattelli-Palmarini (1989) has expressed similar views:
The process of learning just picks them [i.e. innate lexical meanings] out
en bloc, once the phonetic label is assigned to the concept. Association
now becomes a very minor component of the mechanism: What is being
associated is just the sound for that concept and the innately available
concept itself . . . Association is marginal to the process: it boils down
to a heavily constrained pairing of certain labels to certain innate con-
cepts. (Piattelli-Palmarini 1989: 31)
The problem with this conclusion, impeccable as the logic may be, is that it is
patently absurd, since it commits us to the belief that humans are born with
innate concepts such as inning, poppadom, widget, aerobics, Internet, baptise
and overdraft. While such a position is logically consistent, I suspect that
anyone who has ever tried to explain what an inning is to someone with no
experience of baseball or cricket, or what an overdraft is to someone from a
culture that does not use banks, will be rather sceptical.
Furthermore, as Gleitman (1990) convincingly demonstrates, matching the
concept and the phonological label is not as trivial a task as Piatteli-Palmarini
makes it out to be. Determining which aspect of the world a particular part of
a parental utterance refers to is dicult enough even when it refers to some-
thing in the childs immediate environment; but most parental speech is not a
running commentary on situations observed by the child. Utterances such as
Eat your peas, Lets go to the park or It rained last night do not co-occur with
episodes of pea-eating, going to the park, and rainy weather; and episodes of
thinking, knowing, understanding, hoping, wondering, guessing and suppos-
ing all look quite similar. More damaging still, asserting that lexical concepts
are innate hasnt quite eliminated the miraculous: it has merely shifted it to
another domain. We still need to explain how the species acquired these innate
abilities (cf. Chapter 6, section 4.2). A moments reection will reveal that they
could not have come into being through any known evolutionary process
unless thousands of years ago evolution knew that someday we were going
to invent baseball, banks, and tness clubs. Hilary Putnam makes a similar
point:
To have given us an innate stock of notions which includes carburettor,

bureaucrat, quantum potential, etc., as required by Fodors version of
the Innateness Hypothesis, evolution would have had to be able to antic-
ipate all contingencies of future physical and cultural environments.
Words 83
Obviously it didnt and couldnt do this. (Putnam 1988, quoted in

Chomsky 1991: 32)
Chomsky nds this argument unconvincing: evolution didnt need to antic-

ipate all contingencies, only those required for concept formation in the real
world; and we shouldnt worry if the idea seems absurd since many claims
which originally seemed highly implausible have been shown to be true.
Besides, it is misleading to compare ordinary language acquisition with the
acquisition of scientic concepts. The mind has other capacities for the for-
mation of concepts used, for example, during the construction of scientic
theories so concepts that do not fall within the range of [the language
faculty] might be constructible by the mechanisms of [general cognition]
(Chomsky 1991: 33).
So at least some concepts (for example, scientic concepts such as quantum
potential) are not innate and the mind has non-linguistic mechanisms capable
of constructing them. The obvious question to ask at this juncture is which
category peek-a-boo, weekend and table belong to? If Chomsky wishes to claim
that they are innate, then he needs to explain how they could possibly have
been written into the human genome and divine creation seems to be the only
possible explanation. If they are not, then why should other lexical concepts
be unlearnable? There is no evidence whatsoever that such culturally specic
notions are any more dicult to acquire than universal or near-universal con-
cepts as person, three or hot.
Chomskys views on the innateness of lexical concepts are admittedly
extreme; most psycholinguists would maintain that lexical concepts them-
selves are not innate, but their acquisition is tightly constrained by various
cognitive mechanisms which are. I have no quarrel with the claim that lexical
development is constrained by innate factors. What remains unclear, however,
is how tightly constrained it is: that is to say, how much detective work the
child needs to do to establish the meanings of the words she hears around her.
1. The semantics of locative terms
To address this question, we will examine the meanings of words belonging to

a particular domain namely, locative terms and then briey review what
we know about their acquisition. Locative terms, or words like in and behind,
which specify the location of one object (the trajector) with respect to another
object (the landmark), provide a particularly interesting test case for research
concerning the role of innate factors in lexical development. If any lexical con-
cepts are innate, or strongly constrained by innate factors, locative concepts
should be among them. They relate to universal aspects of human experience,
and they have obvious survival value for the species, so it is not dicult to
envisage why evolution might have favoured encoding them in the genome.
Moreover, they would also have been useful to our prehuman ancestors, so
evolution would have had plenty of time to operate.
Locative expressions are also comparatively easy to study because they have
fairly concrete meanings, so there is a large body of research on their mean-
ings and their acquisition. These can be interpreted in the context of research
on infant spatial cognition and infant perceptual processing mechanisms and
their neurophysiological bases.
Moreover, it is possible that childrens knowledge of spatial terms helps them
to learn other aspects of their language. It is well known that spatial terms are
often the starting points of grammaticalisation processes which eventually
produce tense, aspect and case markers, complementizers, comparatives,
morphemes expressing causal relations, and other grammatical functors (see
Casad 1996a; Comrie 1976; Dabrowska 1996; Heine 1997; Rudzka-Ostyn 1985;
Traugott 1978). This suggests a conceptual anity between spatial terms and
various grammatical notions, and in fact some linguists have argued that gram-
matical notions should be interpreted in localistic terms even when there is no
historical evidence that they derive from spatial concepts (Anderson 1971, 1994;
Hjelmslev 1935; Jackendo 1983; Miller 1972, 1974). If such conceptual an-
ities do exist, children could use their knowledge of spatial concepts to get a
handle on the more abstract notions encoded by various grammatical functors.
The conventional wisdom on the acquisition of locative terms is that it
involves learning how to apply them to . . . prior knowledge about space
(Clark 1973: 62; see also Johnston and Slobin 1979; Levine and Carey 1982;
Slobin 1973) that is to say, to prelinguistic concepts of containment, support,
occlusion etc. (In this context, prelinguistic doesnt necessarily mean innate:
the concepts could emerge out of prelinguistic sensorimotor experience.)
There is good supporting evidence for this view:
1. Locative expressions are acquired very early. The rst terms usually appear
at 1617 months, and they have been reported as early as 1214 months
(Choi and Bowerman 1991: 100); by the time the child goes to school, the
system is rmly in place.
2. Children acquiring dierent languages acquire equivalent words in roughly
the same order. The order of acquisition is determined by non-linguistic
knowledge, though acquisition may be delayed if the linguistic form is di-
cult (Johnston and Slobin 1979).
3. Errors in the use of basic spatial terms are extremely rare in spontaneous
speech (Choi and Bowerman 1991; Sinha et al. 1994; Tomasello 1987).
4. Children generalise newly acquired concepts to new contexts very quickly,
often within days of the rst use (Choi and Bowerman 1991; Tomasello
1987).
In spite of this, I will argue that learning locative terms does not consist simply
in matching linguistic labels to pre-existing concepts. The main reason for this
Words 85
is that although we may well have pre-linguistic concepts of containment,

support etc., these do not correspond in any direct way to linguistic expres-
sions such as the English in, on etc. The relevant evidence will be discussed
under three headings: conventional construal, polysemy, and cross-linguistic
variation.
1.1 Conventional construal

Objects in the real world come in all shapes and sizes and can participate in an
indenite number of spatial relationships. On the other hand, the linguistic
resources available to code these relationships are generally quite limited: in
English, about two dozen basic spatial expressions. This means that the same
locative term must be used to refer to a variety of relationships between very
dierent objects, and the t of words to the world is often rather less than
straightforward.
Take the English preposition in. One can dene in as within an inside
region; but what counts as an inside region is anything but obvious. For
example, if the landmark, or reference object, does not dene an enclosed
space, the conceptualiser must project virtual boundaries. With a landmark
such as a bowl, the inside is partly dened by the wall and the bottom, but to
get an enclosed space you need to add a virtual lid, which would normally be
the plane containing the top edge of the bowl. This might seem to be a trivial
point, but consider the congurations depicted in Figure 7.1.
In (a), the banana is said to be in the bowl even though it is not actually
located in the region dened by the walls of the container and the projected
lid; in (b) it is located within this region, but you can only say it is under not
in the bowl. Moreover, in some cases there is more than one way to dene the
Figure 7.1 Is the banana in the bowl? (adapted from Herskovits 1985: 19)
inside region: compare There is a banana in the bowl and There is a crack in the
bowl. Or consider the expression The bird is in the tree: in this case, all the
boundaries of the container are virtual. (All three examples are taken from
Herskovits 1985, 1986.)
Making words t the world also involves a fair amount of idealisation, as
objects need to be assimilated to the abstract geometric descriptions embod-
ied in the semantics of locative expressions. Thus a frying pan might be ideal-
ised as a container (cf. the expression in the frying pan); a river as a line (along
the river) and a table as a horizontal surface (on/under the table) (Dabrowska
1993; Herskovits 1985; Talmy 1983).
The crucial point here is that all of these are conventional, and hence lin-
guistic, phenomena. This can be easily demonstrated by showing that dier-
ent languages require dierent construals. In English, a bird is in the tree, but
many other languages require dierent terms (usually the equivalent of on;
notice that even in English, an apple would be on, not in, the tree, in spite of
the fact that the relationship between the apple and the outer boundary of
the tree is exactly the same as that between the bird and the outer boundary
of the tree). In English, the inside of a bowl can be either the region dened
by the wall, the bottom and the virtual lid (of whatever shape) or the bowl
itself, but many languages allow only the former conceptualisation. Hence, if
you want to say There is a crack in the bowl, you either have to use another
preposition (e.g. the equivalent of on) or a non-spatial paraphrase (something
like The bowl is cracked or The bowl has a crack). Furthermore, dierent lan-
guages require dierent idealisations of the same objects. English idealises a
frying pan as a container, so your fried eggs are in the pan; in Polish, they
would be on the frying pan (na patelni). If a ball rolls between the legs of a
table and emerges on the other side, an English speaker will say that it rolled
under the table: English requires that the table be construed as a horizontal
surface. In contrast, in Atsugewi, the ball would roll through the table because
the table is thought of as a three-dimensional object (Talmy 1983). Sometimes
even dierent dialects of the same language require dierent conventional
construals. In Britain, you leave your car in the street. In America you leave it
on the street, but your neighbours children playing near it would be in the
street; and if you decided that the driveway would be a safer place to park,
your car would in the driveway, not on it.
As a nal illustration, let us consider an extended example from Cora, an
Uto-Aztecan language spoken in a mountainous region of Mexico. Cora has
an extremely rich system for encoding spatial relationships which includes
spatial adverbs, particles, and verbal prexes which are used in various com-
binations to convey elaborate descriptions of spatial congurations. Cora
locatives have been extensively studied by Eugene Casad (1977, 1982, 1988),
and the reader is referred to these works for a comprehensive description.
Because the system is so rich, it cannot be described in any detail here; but its
very richness and utter foreignness are an excellent illustration of the central
Words 87
theme of this chapter, namely, that humans possess complex abstract systems
which are extremely unlikely to come prewired.
The following discussion is based on data from Casad and Langacker
(1985), which describes the semantics of just two morphemes, u (which can be
glossed as inside) and a (outside). Because of space limitations, I cannot
justify the analysis proposed by the authors. However, it should be stressed
that Casad and Langacker do not ask the reader to take the analysis on faith;
as demonstrated throughout the paper, various quirks and idiosyncrasies of
Cora locationals make perfect sense given the construal described here.
Although a means outside, it can also be used to refer to location inside a
shallow container such as a basin or pan (as opposed to, say, a jar, which
requires u). In other words, as far as the u/a distinction is concerned, shallow
containers do not count as containers: any object located near the outside
surface of the landmark is considered outside the landmark.
As in the case of the English in and out, the container with respect to which
things are located can be virtual. Consider the complex expressions mah
away up there to the side in the face of the slope and muh away up there near
the centre line of the face of the slope. Both of these locationals consists of
three morphemes. The rst morpheme, m, is a deictic prex which can be
glossed as at a medial distance from the speaker (in contrast to other pre-
xes meaning near the speaker and far from the speaker). The nal h means
on the face of the slope and contrasts with suxes meaning at the foot of
the slope and at the top of the slope. Between the two come u and a, which
mean inside and outside but inside and outside what? Casad and
Langacker argue that the landmark region is dened relative to the line of
sight:
A conceived situation of some complexity is required [to dene the

boundary]. This situation . . . is sketched in [Figure 7.2]. It involves a
vantage point (VP) somewhere in the area at the foot of the slope; when
no semantic conict results, this vantage point is generally equated with
the position of the speaker. From this vantage point, a viewer (normally
the speaker) is presumed to be looking directly up the face of the slope.
The broken arrow in [Figure 7.2] represents the viewers line of sight in
relation to the slope observe that this line of sight runs from the foot
of the slope, along the face, up to the horizon line between the face and
top, but does not curve to include the region on top. The landmark
enclosure for u/a is dened relative to this line of sight: it includes a
narrow region along either side of the line as well as a restricted area at
the foot of the slope surrounding the vantage point. (Casad and
Langacker 1985: 262)
The following sentences provide an even more striking example of the

importance of conventional construal:
MED
VP
Figure 7.2 mah away up there to the side in the face of the slope (adapted from
Casad and Langacker 1985: 262)
(1) a. u- h- k- tya- puu

inside slope short middle BE:planted upright
Its [the dogs] tail is chopped short.
b. a- h- k- tya- puu
outside slope short middle BE:planted upright
Its [the dogs] tail is chopped short.
Both sentences can be used to describe the same objective situation: namely,
to state that the dogs tail has been lopped o. But what does a short tail have
to do with slopes and containment, and why does the sentence with a mean
the same thing as the sentence with u?
In fact, they dont mean quite the same thing, as they would be appropriate
in dierent contexts. (1a) would be used to describe a situation in which the
dog is viewed from behind, as in Figure 7.3a, while (1b) describes a dog viewed
from the side, as in Figure 7.3b. The motivation for using a and u should now
be apparent. In (1a), the slope facing the observer is the dogs rump, and the
tail is inside the region dened with respect to the observers line of sight. In
(1b), the slope is the dogs side, and the tail sticks out outside the region
dened by the line of sight.2 Notice that this usage requires a (metaphorical)
projection of a term normally used to refer to parts of a mountain onto a body
part, motivated by the topographical similarity of a slope and an animals
rump and conventionalised in the grammar.3 While the Cora u may be said to
expresses containment, containment is very much in the eye of the beholder.
Words 89
Figure 7.3 Another virtual container
1.2 Polysemy
Most locative expressions are massively polysemous. For example, according
to Brugman (1988; see also Lako 1987) the English preposition over has
nearly 100 distinct senses, of which 24 designate various spatial relationships.
A few of these are exemplied in (2) below.
(2) a. There is a lamp over the table. [= above]

b. He spread a cloth over the table. [= on]
c. He jumped over the stream. [= across]
d. He has travelled all over Europe. [= all around]
e. He lives over the hill. [= on the other side of]
Polysemy is a sensible way of dealing with the conicting demands imposed

by the great diversity of spatial relationships to be described and by the fact
that the lexicon must be nite (cf. Geeraerts 1988; Lako 1987; Rosch 1975,
1977). What is relevant for the present discussion is that dierent languages
choose dierent solutions to this problem: that is to say, they carve up the
semantic space in diverse and often incompatible ways. This becomes obvious
when we compare the uses of over exemplied in (2) with the corresponding
locative expressions in another language. Each of these senses of over corre-
sponds to a dierent preposition in Polish (nad, na, przez, po and za, respec-
tively). Moreover, each of the Polish prepositions is also polysemous, and the
distinctions they express cross-cut those encoded in the English system.
For a more dramatic illustration of such cross-linguistic dierences, con-
sider the relationships depicted in Figure 7.4.4 English classies them into
three categories , and and
Figure 7.4 Some in, on, and above relationships: key in a door, coin in a box,
crack in a bowl, bird in a tree, apple on a tree, picture on a wall, plaster (Band-Aid)
on a leg, fly on a window, lamp above a table, lamp on a table
Words 91
codes them by means of three prepositions, in, on and above. Some other lan-
guages (e.g. Greek and Spanish) conate and and code
separately. Kikuyu also has only a two-way distinction, but it con-
ates the prototypical on (as in The lamp is on the table) and above, and
expresses everything else with a generalised locative morpheme. Italian,
Portuguese, Swedish, Indonesian, Mandarin and Finnish all make a three-way
distinction, but each carves up the semantic space in a slightly dierent way.
Dutch and German divide the set of relationships into four categories, and
Japanese into eight.
It is clear, then, that having innate categories corresponding to specic loca-
tive lexemes would be of little use to learners, since the detective work involved
in determining which categories out of the thousands of possible ones were
actually employed in their language would not be much easier than creating
them on the basis of the evidence.
On the other hand, innate concepts need not coincide with specic lexemes.
One could claim, for example, that what is innate is category prototypes
(archetypal congurations around which future categories will grow) or some
more basic dimensions along which lexemes dier (semantic features). I will
discuss the latter possibility later on. The former may initially seem quite
attractive, since there is considerably less variation between languages in the
congurations that they choose for category centres. For example, although
the English preposition on, the Polish na and the French sur all designate
dierent categories, the prototypical sense (a trajector object located on top
of, and supported by, a horizontal landmark) is the same in all three languages.
However, a moments reection will reveal that this theory suers from exactly
the same problems as the more radical view espoused by Chomsky, Fodor and
Piattelli-Palmarini. As the discussion below will make clear, the number of
distinct concepts that would have to be encoded in the genome would still be
intractably large; and the learner would still need to determine category boun-
daries.
1.3 Cross-linguistic variation

The dierences in the ways languages carve up space are enormous.
Depending on the language, locative meanings can be expressed by adposi-
tions, nouns, adverbs, particles, verbs, adjectives, verb axes or noun inec-
tions, or they may be distributed across lexemes belonging to dierent classes
(Sinha and Kuteva 1995). Moreover, spatial meanings are often integrated
with other semantic systems (Brown 1994; Sinha et al. 1994; see also the
Tzeltal data below). Consider, for example, Japanese. Japanese has three
means of conveying locative meanings:
1. Verbs which code path of motion, e.g. hairu go in, deru go out, ireru put
in, dasu take out, agaru go up, oriru go down. Functionally, this is the
most important part of the system, as most locative information is con-

veyed by this means.
2. Relational object-part terms, e.g. naka inside, soto outside, ue top, shita
bottom. For example, tsukue no ue ni means on the desk (lit. at the top
of the desk).
3. Locative particles such as ni, de at, in, o through, from, e to, kara from,
made to. Functionally, this is the least important part of the system, as
the particles are highly polysemous and generally add little or no new infor-
mation to that already conveyed by other elements of the sentence.
This system makes it possible to code aspects of the situation in several parts
of the sentence (often redundantly):
(3) Hon-wa hako-ni haitte imasu.

book-TOP box-AT entering is
The book is in the box.
(4) Hon-wa hako-no naka-ni haitte imasu.
book-TOP box-GEN inside-AT entering is
The book is inside the box.
The rst sentence literally means something like The book, having entered the
box, is (there): the verb hairu conates the idea of motion and the path of
motion (from outside inwards). The semantic contribution of the particle ni
is entirely redundant: it does not add any information to that expressed by
haitte imasu. The same information can be expressed using (4), which is even
more redundant, since the locative information is distributed across three
morphemes: the particle ni at, the relational noun naka inside and the verb
hairu enter.
Other languages have locative systems which dier even more from those
found in the familiar European languages. To appreciate the cross-linguistic
diversity, in this section we will look at a few rather exotic systems: the Mixtec
system of body-part terms, Tzeltal dispositional adjectives, and absolute
systems in several languages.
1.3.1 Metaphorical projection of body-part terms

In Mixtec, an Otomanguean language spoken in Oaxaca in Mexico, locative
relationships are described by using body-part terms (Brugman 1983;
Brugman and Macaulay 1986; Lako 1987). For example, someone who has
just climbed a mountain would be located at mountains head (5). On rela-
tionships are described by using several dierent expressions, depending on
the shape of the landmark object: on the mat would be translated mats face
(6); on the table would be at the tables (animal) back (7); and on a branch
becomes at trees arm (8). (All the Mixtec data cited in this section come from
Brugman 1983 and Brugman and Macaulay 1986.)
Words 93
(5) hizaa-re sini zuku

be+located-3SgM head mountain
Hes at the top of the mountain.
(6) se?e-ri hit nu u` yu

son-1Sg lie face mat
My son is lying on the mat.
(7) hisndee sk mes w

be+on animal+back table that
Its on that table.
(8) ndukoo-ri nda?a ynu

sit-1Sg arm tree
I am sitting on the branch of the tree.
Body-part terms are used to refer not only to the relevant part of the land-
mark object, but also to the neighbouring region in space: at mountains head
could mean over the mountain as well as on top of the mountain. Some of
the terms in the system designate human body parts, some animal body parts,
and some can refer to either, which results in rather strange (to us) patterns of
polysemy. For example, the term ciì` belly can be used to refer to the middle
of the front surface of a vertically oriented object (the region where the belly
would be in an upright human) or the underside of a horizontally oriented
object (where the belly would be in a four-legged animal), or the vicinity of
either of these. Thus, with a horizontal landmark such as a table, ciì` can be
translated as under (9).
(9) yu w hiya ciì` mes

stone DET be+located belly table
The stone is under the table.
It should be obvious from this brief description that English and Mixtec
carve up spatial scenes in very dierent ways, and, consequently, the informa-
tion conveyed by a Mixtec sentence and its English translation is not strictly
equivalent. English is quite explicit about the position of the trajector with
respect to the landmark (contact/no contact) but does not specify the shape of
the landmark, while the opposite is true in Mixtec.
On second thoughts, the use of body-part terms to express locative mean-
ings is not as exotic as it might at rst seem. The American English expression
in back of is also a metaphorical extension of a body-part term, and the same
metaphor is (just about) accessible, though much less transparent, in the prep-
osition behind. The word front in the complex preposition in front of comes
from Latin frons forehead, though in this case the link is purely etymological
and unknown to most speakers. In fact, body-part terms are common sources
of locative expressions (see Svorou 1994 and the extensive literature cited
there). What is special about Mixtec is that the use of body-part terms is much
more systematic and more transparent synchronically.
However, the extension of body-part terminology to the locative domain is
conventional: that is to say, although there are broad similarities between lan-
guages in the way they eect the mapping, the details are language-specic and
must be learned. For example, the words for eye, face, forehead, head, breast
and chest are all common sources for the spatial meaning of in front of
(Svorou 1994); but dierent languages make dierent choices from this list. In
this connection, it is also instructive to compare Mixtec and Cora: while
Mixtec projects body-part terms onto inanimate objects such as mountains,5
Cora sometimes projects hilly terms onto parts of the body.
1.3.2 Dispositional adjectives

Like Mixtec body-part terms, dispositional adjectives in Tzeltal code more
than the position of one object with respect to another (Brown 1994). Tzeltal
does not have words corresponding to the English on or in, and information
about and is always co-lexicalised with much more spe-
cic information about shape, size, orientation, mode of support or attach-
ment, subject category, and other semantic dimensions in dispositional
adjectives which can be translated by such complex English expressions as
mounted on top of, positioned at right angles to, hanging down from,
glued to the surface of, spirally wound around etc. There are several
hundred roots which encode very specic shape information about location.
Some examples of these are presented in Figures 7.5 and 7.6.
Tzeltal dispositional terms actually designate the overall conguration of
the trajector and the landmark rather than the relationship between the two;
information about the locative relationship itself is usually conveyed by impli-
cation rather than directly. This leads to interesting ambiguities, which Brown
calls gure/ground ambiguities, deriving from whether the adjective is taken
as describing the gure (i.e. trajector) or the ground (the landmark; see Figure
7.7). For example, waxal ta X Y standing at X Y can mean either (vertically
aligned) Y is on X or Y is in a (vertically aligned) X. The dispositional adjec-
tive merely indicates that something is in a vertical position; it is the following
NPs which specify which of the two objects is the gure (trajector) and which
is the ground (landmark). From this we can infer whether the gure is on or
in the ground.
1.3.3 Absolute systems

The nal example in this brief survey of locative systems in the worlds lan-
guages concerns relationships between objects located in the horizontal plane.
English uses four contrasting terms to designate such relationships (in front of,
behind, to the left of, to the right of ), as well as more general terms like near
and beside. Systems such as this are called relative because, when the landmark
Words 95
Figure 7.5 Some ways of encoding on relationships in Tzeltal

Figure 7.6 Some ways of encoding in relationships in Tzeltal
object does not have an inherent front, the position of the trajector is
described relative to the location of an observer, usually the speaker. Thus,
depending on where the observer is, the same situation can be truthfully
described by saying The mouse is in front of/behind/to the left of/to the right of
the sh (cf. Figure 7.8).
English also uses an absolute system of geographical terms (north, south,
east and west) to describe macrolocations, or locations of large gures such as
towns and villages with respect to large reference objects. When I say
Doncaster is east of Sheeld, I am using the absolute system.
Some languages use an absolute system to describe microlocations as well.
In such languages, the conguration represented in Figure 7.8 will be
Words 97
Figure 7.7 Figure/ground ambiguities

Figure 7.8 A spatial scene and four observers
described by saying something like The mouse is east of the sh. Note that
because the coordinates of the system are xed rather then being projected
from the speaker, the mouse is always to the east of the sh, regardless of the
position of the speaker/observer.
Absolute systems sometimes use the four compass points, and sometimes
other cardinal directions dened by the direction of the prevailing winds
(Guugu Yimidhirr) or some feature of local topology such as a large river
(Yupno) or the sea (many Oceanian languages). In Tzeltal, the relevant topo-
logical feature is the general inclination of the land, which slopes down
towards the north, so anything south of the reference object is said to be
uphill and anything towards the north is downhill. (This is just the overall
direction of the inclination: there are many local irregularities, but the general
direction of the slope xes the system of coordinates. See Brown and Levinson
1993 for further details.) These two opposing directions dene the strong axis
of the system. The system also has a weak axis perpendicular to the strong
axis which lacks opposing poles. Together, the two axes dene a three-way dis-
Words 99
tinction: up v. down v. across the slope. In Figure 7.9, the saucer is uphill, the
spoon is downhill and the two coasters are both across the slope from the
glass.
It should be clear from the above discussion that absolute systems are not
only quite dierent from relative systems; they are also dierent from each
other in that they use dierent reference points and dierent numbers of car-
dinal directions.6 While east and west always refer to the same direction,
inland and seawards do not: if you live on an island, inland means east
on the western coast and west on the eastern coast.
1.3.4 Interim conclusion

In spite of the fact that spatial conceptualisation is strongly constrained by the
nature of the world and by our psychobiology, there is tremendous variation
in the way that dierent languages structure space. This makes the hypothesis
that locative concepts are innate extremely implausible.
A less extreme version of nativism in lexical semantics is the claim that all
humans are born with a set of semantic primitives (features, atoms etc.)
which can be combined in various ways to construct the meanings of all lexical
concepts in all the worlds languages. This approach suers from most of the
problems that plague the holistic approach, plus some of its own, and is thus
even less attractive. Some of these problems will be discussed later, in section
3.1 of this chapter; for the present, we may just note that for our purposes,
most versions of the atomistic approach are indistinguishable from holism
because the primitives postulated by the atomists correspond to lexical cate-
gories in familiar European languages: in corresponds to the semantic primi-
tive , or the notion of , on expresses or , and so on.
down
up
Figure 7.9 The Tzeltal geocentric system of projective relationships
For example, Jackendo (1983) notes that the English preposition under can
designate a place (The mouse is under the table), a path (The mouse ran under
the table The mouse traversed the region under the table) or the endpoint of
a path (The mouse ran under the table The mouse ran to the region under the
table), and proposes the following conceptual structures to capture the dis-
tinctions:
(10) a. [Place ([Thing ])]

b. [Path ([Place ([Thing ])])]
c. [Path ([Place ([Thing ])])]
The expression up the hill (as in The house is up the hill) receives the following
analysis:
(11) [Place ([Path ([Thing ])])]
Thus, according to Jackendo, , and are semantic primitives

whose meanings correspond to the basic meanings of the English prepositions
on, up and under. Now there is nothing incoherent about the claim that the
meaning of a particular expression is paraphrasable by means of a single
semantic primitive. However, as Bowerman points out, there are reasons to
worry about such glib formulations:
I nd it sobering that the non-linguistic spatial concepts often hypothe-

sised to underlie spatial prepositions e.g. containment and support
lend themselves much more readily to shaping into spatial categories
of English than, say, Tzeltal. In other words, our ideas about plausible
primitives in the language of thought may themselves be conditioned
by the language we have learned. (Bowerman 1996: 160)
2. The acquisition of locative terms
As explained earlier, the conventional wisdom on the acquisition of locative

terms is that it involves matching linguistic labels to pre-existing concepts.
Now that we have seen the diversity in locative systems across languages, this
position is rather less appealing: it is extremely unlikely that children are
equipped with innate concepts corresponding to locative terms in all the
worlds languages, and even if they were, such a vast inventory of possible
meanings would be of little help in xing the meanings of the words which
actually occur in the childs language.
However, it is possible, in principle, that children come equipped with basic
locative concepts which happen to correspond rather closely to the meanings
of English prepositions. If this were the case, then locative terms which do not
Words 101
correspond to the innate concepts should be considerably more dicult to

acquire. In this section, therefore, we briey examine some research devoted
to the acquisition of locative systems which structure space in radically dier-
ent ways from English and other familiar Indo-European languages.
2.1 Warlpiri
We will begin by looking at the acquisition of locative terms by Warlpiri chil-
dren. Warlpiri is interesting from our point of view because it does not have
terms corresponding to in, on and under. It does, of course, have terms which
can be used to talk about scenes which in English would be described using
these prepositions, but these are extended senses of expressions whose
primary meanings are rather dierent. The terms are kanunju-mparra down-
along or, in certain contexts, under; kaninja-rni down-hither or, with the
right kind of landmark, in; and kankarla-rra up-thither or on. All three
expressions are bimorphemic and consist of a locative stem and a directional
ax. It is not very clear what the exact semantic contribution of each mor-
pheme is the traditional glosses are not very helpful in this regard but it is
evident that Warlpiri spatial categories are dierent from those found in
English and other familiar languages.
Bavin (1990) conducted a simple experiment testing the comprehension of
locative terms by Warlpiri children aged from 3;11 to 8. The experimental task
involved placing an object as instructed by the experimenter in, on or under
a landmark object. The landmarks used were a toy yunta (a shelter consisting
of a at top with four poles for support), a toy yujuku (a shelter with sides,
open at the two ends), and a toy pick-up truck.
The results were quite intriguing. Kankarla-rra up-thither was fairly con-
sistently interpreted as on by all the children, but responses to instructions
with the other two terms were varied, and depended on properties of the ref-
erence object. Kanunju-mparra down-along was interpreted as under when
the reference object was a yunta; but when the landmark was a truck, 39 per
cent interpreted it to mean in the open back of the truck. Kaninja-rni down-
hither was interpreted as in when the landmark was a yujuku, but when the
landmark was a truck, the children gave inconsistent responses: some placed
the object in the cabin, some on the tray in the back and some under the truck.
In other words, there is no evidence in Bavins data that Warlpiri children are
sieving through the language input looking for labels for their prelinguistic
concepts , and . In fact there is no evidence that they make these
distinctions at all.7
2.2 Tzotzil
Tzotzil is a Mayan language, closely related to Tzeltal. Like Tzeltal, it uses geo-
centric coordinates dened by the general incline of the land, but in the region
where Tzotzil is spoken the land slopes westwards, so uphill means east, not
south.
De Len (1994) found that Tzotzil caregivers use geocentric terms
(uphill/east and downhill/west) with children as young as 2, and children are
expected to understand the terms. (One mother in the study rebuked her
3-year-old son for not knowing where east was!) Children know the basic prin-
ciples of the system by around 4;6 (i.e. before European children reliably dis-
tinguish left and right), though they generally use local landmarks rather than
geocentric terms: that is, instead of saying The basket is uphill/east, they
might say something like The basket is towards the market town. They start
producing the terms for sunset and sunrise around 9, which suggests that
around this age the child shifts the local centre of the orientation grid to the
absolute co-ordinate provided by the sun (de Len 1994: 876).
Like Mixtec, Tzotzil also uses body-part terms to describe more immediate
locations. Preliminary work carried out by de Len (1991, cited in Levinson
1994) suggests that this system is acquired even earlier: Tzotzil children begin
to use body-part locatives well before the age of 3.
2.3 Japanese
In Japanese, we saw earlier, locative meanings are distributed across three form
classes: motion verbs, relational nouns and particles, with the rst of these
doing most of the work. Sinha et al. (1994) report on a longitudinal study of
one Japanese child, Adam, observed by the team from age 1;3 to 2;11.
Adams rst locative terms were motion verbs such as dasu take out, toru
take o, hameru t in etc. The rst recorded usage of a locative verb
occurred at 1;3.28; and they are recorded in every session from 1;6 onwards.
Locative nouns (soto outside, naka inside, ue top etc.) emerged shortly
afterwards, around the time of his second birthday. The last to appear were
the particles ni, de at, in, kara from, and e to. All four particles rst
emerged in collocation with koko this place, and their use remained highly
formulaic throughout the period of study. This pattern of acquisition makes
sense, since it is motion verbs and relational nouns that convey specic infor-
mation but it is rather dicult to explain if you believe that children are
simply looking for labels for pre-existing locative concepts corresponding to
in, on etc. If Adam were simply looking for labels for in and on, he should
have learned the locative particles much earlier, since they occurred fairly fre-
quently in the input (in fact, the frequency of locative particles in the speech
of Adams mother was four times that of locative relational nouns, which he
acquired earlier).
This acquisition pattern, incidentally, also argues against the idea that chil-
dren construct meanings out of a universal set of primitives. If this were the
case, more specic concepts should be more dicult to acquire than concepts
which directly correspond to the putative primitives. The data collected by
Words 103
Sinha and his colleagues suggests the opposite: the verb hameru t in, which
co-lexicalises the action of inserting and the destination of the patient, is
acquired earlier than the general locative particle ni at, in; toru take o,
which co-lexicalises causative movement and direction, comes long before
kara from, and so on.
2.4 Korean
Another series of studies conducted by Melissa Bowerman and Soonja Choi
(Bowerman 1996; Choi and Bowerman 1991; Choi et al. 1999) compared the
acquisition of morphemes specifying the path of a moving object by English-
and Korean-speaking children. The relevant parts of the locative systems of
the two languages are quite dierent. English path morphemes are preposi-
tions, while in Korean the shape of the path is co-lexicalised with other notions
in the verb. Furthermore, English path morphemes are indierent as to
whether the movement is spontaneous or caused (the same preposition is used
with both transitive and intransitive verbs: go in, put in), while in Korean the
corresponding meanings are expressed by dierent verbs. Finally, Korean path-
incorporating verbs make distinctions which cross-cut those made by English
prepositions: e.g. kkita t a three-dimensional object into or onto another
object can be translated as put in (for example, when describing the action of
putting glasses inside a case), put on (when a cap is put on a pen), or put together
(when two Lego pieces are tted together). Conversely, the meanings expressed
by put in and put on are coded by a number of dierent verbs in Korean: kkita
when two three-dimensional objects t closely, nehta when the t is loose, pwu-
thita when two at surfaces are joined, kkotta when an elongated object is put
into/onto a base, and so on. In addition, there is another set of verbs for des-
ignating the action of putting clothing on various parts of the body.
Choi and Bowerman report that the rst locative expressions usually appear
between 14 and 16 months, and that children use them productively by 1620
months. The childrens use of locative expressions is generally correct, and it
follows language-specic patterns from the very beginning. English-speaking
children acquire their rst path morphemes (up, down, in, out, on, o ) very
early (usually between 16 and 17 months, and sometimes as early as 1214
months) and almost immediately generalise them to spontaneous and caused
motion and to changes of body posture. For example, they might say in when
putting something into a box or when they themselves are climbing into a box;
up when they want to be picked up, when going up stairs, and when standing
up. Korean children, in contrast, always use dierent verbs to refer to caused
motion and to spontaneous motion. Thus, they will say kkita when tting an
object into a hole, but they never overextend this term to describe a situation
when they themselves creep into a narrow space; and they never overgeneral-
ise ancta sit down and anta pick up and carry to situations involving upward
or downward motion.
Furthermore, English-speaking children distinguish (in) and

(on) relations, and do not worry about details of trajector and land-
mark geometry, or whether the t is tight or loose. Korean children do not
make this distinction: they use kkita t for putting things into tight contain-
ers and for tight attachment to a ground object, nehta for loose containment,
kkota for putting an elongated object into/onto a base, and so on.
In short, Bowerman and Choi found no evidence that children begin by
mapping prelinguistic spatial concepts onto the words they hear. In spite of
this, the children in their study made few errors. When errors did occur, they
tended to be language-specic, showing that the children had diculty with
working out the details of a system they had already acquired. Korean chil-
dren, for example, had problems with words that encode pure path notions,
which is an English lexicalisation pattern. This shows that even at this very
young age (under 2), they had already formed some generalisations about
what was likely to be lexicalised in their language and what was not.
2.5 General discussion

The developmental studies summarised here clearly show that children
acquire the rudiments of the locative system of their language very early well
before the age of 3, though some distinctions may take somewhat longer to
learn. Moreover, they appear to be sensitive to language-specic categories
from the very beginning, and the pattern of acquisition observed in naturalis-
tic studies is virtually error-free (Choi and Bowerman 1991; Sinha et al. 1994;
Tomasello 1987). The developmental research, then, shows that children are
extremely good learners.
I observed at the beginning of this chapter that several studies (Carey 1978;
Dickinson 1984; Heibeck and Markman 1987) have shown that children can
form partial semantic representations of new words after only a few expo-
sures, or even just one exposure, to the new word. These fast-mapping studies
have been used as evidence that lexical knowledge must be innate (though the
authors themselves did not necessarily advocate such conclusions). However,
the existence of fast mapping does not necessarily imply that children are
endowed with innate lexical knowledge. It could also be interpreted as further
evidence that they are good learners. This conclusion appears to be conrmed
by a study by Markson and Bloom (1997), which showed that fast mapping is
not conned to word learning.
Markson and Bloom asked their subjects (3- and 4-year-old children and
adults) to use unfamiliar objects to measure the length of another object. In
one version of the task, the objects were referred to as kobas (Lets use the
kobas to measure which is longer. Line up the kobas so we can count them.
We can put the kobas away now.). In another version, participants were
simply told that the objects were given to the experimenter by her uncle (We
can use the things my uncle gave me to measure which is longer. My uncle gave
Words 105
these to me. We can put the things my uncle gave me away now.). This was fol-
lowed by a testing phase, which occurred immediately, after a weeks delay, or
a month later. In the testing phase, participants were shown an array of objects
and were asked to pick out the koba or the object given to the experimenter
by her uncle.
Markson and Bloom found that adults performed signicantly better than
children on the word-learning task when they were tested immediately after
training, but the dierences between age groups disappeared after a week.
About 70 per cent of the participants could still point to the correct object
when tested a month after training. This replicates the results obtained in
earlier fast-mapping studies. The new nding was that participants were
equally good at remembering which object was given to the experimenter by
her uncle; in fact, when tested immediately after the training phase, children
performed signicantly better on this task than on the word-learning task.8
Fast mapping, then, is not restricted to word learning.
On the other hand, Markson and Bloom also found that people are not
equally good at remembering any arbitrary fact about an object. In another
experiment, participants watched as the investigator placed a sticker on one of
the unfamiliar objects and heard a similar commentary (Watch where this
goes. This goes here [placing sticker on one of the objects]. Thats where this
goes.). In this version of the experiment, recall was signicantly poorer. It is
not clear whether this was due to the fact that the information was presented
visually or whether the location of a sticker was deemed less relevant by the
participants than the identity of an object. Whatever the reason, their ndings
suggest that fast mapping, though not unique to language, does not apply to
all memorisation tasks. Children are good learners, but not equally good at
learning anything. Something about the human cognitive apparatus makes
certain associations easier to acquire than others.
3. Innate structure, yes but of what kind?
The general conclusion, then, is that children are good learners, but they are
not unbiased learners. In fact, many would argue that they are good learners
because they are biased towards certain solutions. It is important to realise that
rejecting the notion that lexical concepts are built in does not entail rejecting
the existence of innate structure of any kind. Clearly, children must have some
innate predispositions which make language learning possible. Learning is
impossible without experience, but an organism needs a great deal of innate
structure before it can have any experience, and even more to be able to extract
patterns from it. There is considerable empirical evidence for various con-
straints on lexical development in general (Dockrell and Campbell 1986;
Markman 1987, 1990; Smith et al. 1992; Tager-Flusberg 1986) and for prelin-
guistic biases on the perception of space which shape the acquisition of
locative terms (E. V. Clark 1973a, 1976, 1980; H. Clark 1973; Huttenlocher et
al. 1983). The question, then, is not whether children are endowed with innate
structure, but rather what form their innate predispositions take, and it is to
this question that we now turn.
3.1 Semantic primitives

A less extreme version of nativism in semantics is the claim that what is innate
is not lexical concepts themselves, but semantic primitives or atoms of
meaning like , or . These can be combined to form complex
concepts such as [ ] and [-
--] (see Bierwisch 1970; Bierwisch and Schreuder 1992;
Jackendo 1983, 1990; Katz 1972, 1987; Katz and Fodor 1963; Miller and
Johnson-Laird 1976). On this view, lexical acquisition involves discovering
which combinations of primitives are lexicalised in a given language (Clark
1973b, 1975; Clark and Clark 1977).
Although the approach may initially seem quite attractive, it is plagued with
problems, many of them of a rather fundamental nature. Perhaps the most
serious of these is that most lexical concepts cannot be analysed into primi-
tives. Semantic decomposition works reasonably well for a few lexical elds
such as kinship terms and farmyard animals, but attempts to extend it beyond
these have been unsuccessful. As noted by Fodor et al. (1980), the number of
convincing exhaustive decompositions in the literature is vanishingly small.
Most decompositions are partial: some aspects of meaning are captured by
means of semantic primitives, but when these are factored out, we are left with
an unanalysed residue. Take the word red, for example. Red is a colour term,
so we could say that its semantic representation contains the feature [].
Let us factor this out of the meaning of red. What do we do with the residue?
Is it coherent to talk of redness minus colour (cf. Fodor et al. 1980; Jackendo
1983)?
There are problems even with traditional mainstays of semantic decompo-
sition. As several researchers have pointed out, kill is not synonymous with
cause to become not alive or even cause to die: you can cause someone to
die on Tuesday by shooting him on Monday, but you cannot kill someone on
Tuesday by shooting him on Monday (Fodor 1970; Lako 1987). In a similar
vein, there are various connotational dierences between the terms bachelor
and spinster which are not captured by the traditional feature denitions
(bachelor: [ ]; spinster [
]): spinsters are supposed to be old, sexually unattractive and
desperate to get married, while bachelors are either young and eligible or
happily single and not interested in marriage.9 Worse still, some unmarried
adult males for example, priests, homosexuals, and male partners in cohab-
iting heterosexual couples cannot be felicitously described as bachelors,
except in a strictly legal context (Lako 1987: 70).
Words 107
Another diculty concerns the meaning of the features themselves. As

many critics have pointed out, proponents of semantic decomposition have
failed to provide a satisfactory account of how speakers translate perceptual
information into semantic primitives (see Johnson 1987; Lako 1987); and
attempting to dene words by using meaningless features is a rather pointless
exercise.
Semantic feature theory also has little to recommend itself in the way of
psychological reality. Intuitively, concepts such as and are, if
anything, more basic than the supposed primitives and .10 Such
intuitions are corroborated by psycholinguistic evidence. A number of experi-
mental studies attempted to determine whether lexical decomposition is a
prerequisite to comprehension by comparing reaction times to sentences con-
taining words expressing simple concepts (e.g. ) with those containing
more complex concepts such as or (both of which refer to a particu-
lar kind of motion). If comprehension involves lexical decomposition, the
latter should take longer to process. Most studies, however, have failed to nd
any dierences, and some have found the reverse of the expected result: reac-
tion times to complex concepts were shorter than to concepts dened using
fewer features (Bierwisch and Schreuder 1992; Fodor et al. 1975).
Developmental studies also fail to provide support for the notion of seman-
tic primitives. As pointed out by Taylor (1989), the semantic feature hypothe-
sis predicts initial overextension followed by contraction, in discrete steps, of
category boundaries (as the learner adds new features); the categories the child
uses during the overextension phase should be similar to adult superordinate
categories. Neither of these predictions is conrmed by developmental data.
Early usage is often characterised by underextension, which is followed by a
period of gradual extension and contraction during which usage at category
boundaries oscillates while the category centre remains stable. This develop-
mental pattern is consistent with the predictions of prototype theory but not
of semantic feature theory. Furthermore, childrens rst words rarely express
primitive notions. On reection, this is hardly surprising, since complex
notions those expressing meanings dened by a combination of features
tend to be more concrete and hence evoke richer imagery (cf. move and y, tool
and hammer, container and bottle).
3.2 Perceptual primitives

The semantic primitives approach, then, though it may initially seem quite
promising, turns out to be unworkable. This brings us back to the question of
how we come to have concepts in general, and spatial concepts in particular.
If they are not innate, they must be learned from experience but how? The
traditional empiricist answer is that complex concepts are constructed out of
simpler concepts which ultimately derive from sensorimotor experience. Is this
solution at all feasible?
Most cognitive scientists would argue that it is not. The following quotation
from Fodor et al. (1980) sums up the prevailing views on the matter quite well:
Its thus important to emphasize that the Empiricist version of TSP [The
Standard Picture, i.e. semantic decomposition theory] is not plausible. If
there are few convincing examples of denitions, there are literally no
convincing examples of denitions which take prima facie non-
sensory/motor terms into a sensory/motor vocabulary. There is, indeed,
no reason to believe even that the denitions we have examples of gen-
erally exhibit an epistemically interesting direction of analysis. On the
contrary, the right hand (deniens) side of standard examples of de-
nitions does not tend, by and large, to be more sensory/motor than the
left hand (deniendum) side. So, for example, the conventional wisdom
has it that grandmother means mother of father or mother; ketch
means two-masted sailing vessel such that the aft (or mizzen) mast is
stepped forward of the helm; bachelor means unmarried man; kill
means cause to die, etc. Its hard to discern a tendency toward the
sensory/motor in these examples, and these examples are quite typical.
Surely, no one could seriously argue that words like mother, father,
mast, vessel, cause, die, married and man, all of which occur in the
deniens, are somehow closer to transducer language, than, say, grand-
mother, ketch, kill or bachelor are. (Fodor et al. 1980: 268)
Fodor et al. are right, of course, when they argue that, for the overwhelming
majority of lexical concepts, verbal denitions are no nearer to sensorimotor
experience than the terms dened. However, there is no reason to think that
the semantic representations of lexical concepts are verbal denitions. Let us
take one of their own examples: the meaning of ketch. While it is true that
terms such as mast or vessel are no more perceptual than the term dened, it
does not follow that our mental representation of the concept ketch does not
include a perceptual component. In fact, I would argue that a visual image is
a central part of the meaning of the term. One acquires the concept that goes
with the phonological label ketch by seeing exemplars and extracting their
shared properties. A verbal denition is helpful to the extent that it draws
attention to the relevant characteristics, but it is neither necessary (since it is
possible to acquire the concept without ever hearing the denition) nor su-
cient (a person who had been presented with a denition but never seen the
real thing would have an underspecied representation of the concept).
One might argue that knowing the meaning of ketch involves much more
than knowing what it looks like: equally relevant are the knowledge that it is
a type of ship (and hence that it is used for travelling over water), an under-
standing of how it moves, and so on. While these observations are undeniably
true, they would only constitute a valid objection if this kind of knowledge
could not be represented using symbols derived from perceptual experience.
Words 109
Up until the mid-1980s, most cognitive scientists regarded it as axiomatic that

indeed it couldnt: perceptual representations were considered as no more than
attenuated copies of sensory states produced by external stimulation, and
hence in principle incapable of representing abstract concepts of any kind.
However, recent work in cognitive linguistics (Johnson 1987; Lako 1987;
Lako and Johnson 1980; Langacker 1987, 1991a, 1991b; Rudzka-Ostyn
1988; Talmy 1983, 1985), cognitive psychology (Barsalou 1993, 1999;
Barsalou et al. 1993; Gibbs and Colston 1995; Jones and Smith 1993; Mandler
1992; Paivio 1986), and articial intelligence (Regier 1996) shows that symbol
systems grounded in perceptual experience are much more powerful than pre-
viously believed. Their new power derives from shedding two traditional
assumptions about sensory experience: the ideas that perceptual images are
mental pictures, and that they are unstructured holistic representations.
The impetus for the new work came from a number of studies of perception
which showed that certain neurons in the brain respond selectively to dierent
aspects of the stimuli perceived. For example, some neurons in the visual
cortex respond to the presence of a line irrespective of its length, orientation
or position; other neurons re only in response to stimuli with a specic orien-
tation; and so on. If dierent aspects of a visual image (shape, orientation,
colour, movement etc.) are processed separately during perception, there is no
reason to assume that they must be accessed together during conception. This
implies that perceptual images can be partly indeterminate, and that they are
componential.
Barsalou (1999) provides an excellent synthesis of much of the new work
on perceptual symbol systems and argues that knowledge derived from per-
ception can support a fully blown conceptual system with the capacity to rep-
resent abstract concepts, combine concepts productively, and produce
inferences. In his framework, a perceptual symbol is a record of the neural
activity that occurs during perception. Patterns of activation in the sensori-
motor areas of the brain are captured in the association areas during percep-
tion. In conception, the association areas partially reactivate sensorimotor
areas to simulate the sensorimotor experience. Thus, concepts are mental sim-
ulators which can give rise to slightly dierent simulations on dierent occa-
sions.
As suggested earlier, these new perceptual symbols are componential (dier-
ent aspects of concepts are represented separately) and schematic (indetermi-
nate with respect to some properties). Another important characteristic is that
they are multimodal: they combine information from the various sensory
modalities, proprioception and introspection. Introspective information (i.e.
information about the conceptualisers own mental states) is particularly
important, as it provides the foundation for the development of abstract con-
cepts. Consider, for example, Barsalous account of how the concept
could be implemented in a perceptual symbol system. To represent the fact that
a particular statement is true, the conceptualiser produces a simulation of the
event that it describes and compares it with his own perception of the event; if
the two structures map onto each other, the sentence can be described as true.
While performing the comparison, the conceptualiser develops perceptual
symbols for the simulation of the event, the process of comparing the simula-
tion to the perceived scene, and the outcome of the comparison (a successful
mapping). If the process is repeated on a number of occasions, a simulator will
develop for the concept : that is to say, the conceptualiser will have
learned to simulate the experience of successfully mapping a simulation onto
a perceived situation (Barsalou 1999).
This proposal, though admittedly sketchy and speculative, shows that it is
possible to represent abstract concepts using experientially grounded symbols,
at least in principle, and hence the pessimism expressed by Fodor and others
is unfounded. A dierent, but not necessarily incompatible approach to
abstract concepts is to be found in the work of many cognitive linguists, who
have argued that these concepts are understood at least partly in metaphori-
cal terms, that it to say, by importing structure from more basic domains. For
example, it has been suggested that is understood partially through the
metaphor : hence, when people
get angry, we say that they get hot under the collar, that their blood boils, that
they get all steamed up, and eventually, they may explode (Kvecses 1986;
Lako 1987; Lako and Johnson 1980; Matsuki 1995). Other linguists have
proposed that mental processes such as understanding and reasoning are
understood in terms of more accessible concepts such as and
(see Jkel 1995; Lako and Johnson 1980).
It seems, then, that the idea that meanings are built up from perceptual
primitives does oer a feasible solution to the problem of how we come to have
concepts. Clearly, many problems remain to be solved and many important
details still need to be elaborated. However, the considerable progress that has
already been made suggests that this is a very promising avenue of research. It
should also be emphasised that, other things being equal, there are several
reasons for preferring theories which use perceptual primitives over those
which appeal to amodal systems. First, using perceptual primitives solves the
symbol-grounding problem. Theories which postulate amodal symbols must
provide an account of how these are mapped onto perceptual states and enti-
ties in the world.11 This problem never arises in a system which uses percep-
tual symbols, since these are built bottom-up from perceptual states. Secondly,
using perceptual primitives dispenses with the need to postulate biologically
implausible innate structures with no obvious evolutionary antecedents.
Perceptual systems have been around for a long time and the neurological
structures that they use provide an obvious source of building blocks for
higher-order concepts. Last but not least, perceptual primitives, while con-
straining learning in highly specic ways, oer enough exibility to support
the wide variety of lexical concepts that we encounter in the worlds languages.
All this said, the proof of the pudding is in the eating, and a tting way to
Words 111
conclude this chapter is to ask whether a system equipped only with percep-
tual structures and a general learning mechanism could acquire the locative
terms of the various languages that have been discussed earlier. The ultimate
test of the proposal would be to design a computer model which had the basic
perceptual structures built in and see how well it would perform on spatial cat-
egorisation tasks. As it happens, such a model has been built, thus oering us
an opportunity to evaluate the proposal.
4. Lexical learning in a constrained connectionist network
The model alluded to was developed by Regier (1996). Regier used an

approach he calls constrained connectionism, which diers from ordinary
unconstrained (i.e. parallel distributed processing) approaches in that it con-
tains structures dedicated to specic tasks. The structures bias the learning
process, making some things easy and others very dicult or impossible to
learn. However, the constraints are perceptual, not linguistic; and unlike the
innate structures postulated by generative linguists, they are independently
motivated.
Regiers model contains three built-in structures: an orientation-combi-
nation structure, which extracts various orientational features; a map-
comparison structure, which detects quasi-topological features such as
contact and inclusion; and a motion-trajectory structure, which parses events
into source, path and goal and calculates the overall path of the trajector. The
rst two structures are loosely motivated by their functional similarity to
known neural structures (orientation-sensitive cells, topographic maps of the
visual eld, cells with centre-surround receptive elds); the motivation for the
motion trajectory structure is less direct (see Regier 1996). All three structures
are embedded in a traditional connectionist network.
Regier presented the network with movies consisting of sequences of tra-
jectorlandmark congurations. Some of these depicted stationary congura-
tions; in others, the trajector moved relative to the landmark. Each movie was
paired with a linguistic label (in, on, through etc.). The network was trained
using English, German, Russian, Japanese and Mixtec terms (on dierent
runs) and then tested on a novel set of movies.
The results were very encouraging. The network was able to learn to
respond with the appropriate terms in all ve languages and showed excellent
generalisation to novel congurations. As we have seen earlier, the languages
used in the experiment carve up the semantic space in dierent ways, and
hence the network had to learn to attend to dierent features of the input,
depending on the language. For example, consider the terms for above and
on relationships. In English, the relevant feature is contact: if the trajector
touches the landmark, it is on it; otherwise, it is above the landmark. In Mixtec,
the contrast hinges on the orientation of the landmark: sini head is used when
the landmark is vertical and sk animal back when it is horizontal, regard-
less of whether there is contact or not. German requires its speakers to pay
attention to both features: auf is used when the trajector is in contact with a
horizontal surface, an when it is in contact with a vertical surface, and ber
when there is no contact. Finally, Japanese has only one term, ue ni, which is
insensitive to either distinction. The network, like a human child, learned
whichever system it was exposed to. The built-in perceptual structures oered
enough constraints to make learning possible and enough exibility to allow
learning systems based on dierent principles.
It should be emphasised that Regiers network is not a realistic model of
human spatial cognition. The structures built into the model are only loosely
motivated by known neural structures, the relationship being one of analogy
rather than exact correspondence. The network did not have to solve all the
complexities discussed earlier in this chapter: its input was highly idealised
(the objects in the movies were simple, two-dimensional gures labelled tra-
jector and landmark) and it did not have to deal with the additional compli-
cations posed by polysemy or virtual boundaries. Furthermore, the network
lacked the prelinguistic knowledge that we know children bring to the acqui-
sition task, and it had no social experience (it did not interact with other
agents). However, the fact that it was able to extract the relevant features of
spatial congurations and succeeded in generalising them to novel instances
shows that it is possible to build semantic representations from perceptual
primitives, and thus oers indirect support for the conjecture that this is what
children do as well.
5. Conclusion
A common argument for the innateness of syntax is that children could not
learn certain aspects of the grammar of their language through any known
learning process, and therefore at least some syntactic knowledge must be
innate. This is an example of what Dawkins (1986) calls an argument from
incredulity: I dont see how children could possibly learn syntax, and therefore
they cannot learn it it must be innate.
Explaining language acquisition by appealing to innate knowledge does
not answer the question of how we came to have language; it just shifts it from
an ontogenetic to a phylogenetic level. Saying that the species acquired lan-
guage through some unknown evolutionary process is not an advance on
saying that children acquire language through some unknown learning
process. The unlearnability argument was generally accepted partly because
the generative tradition in psycholinguistics has tended to overestimate chil-
drens grammatical accomplishments (that is to say, impute to them knowl-
edge that they do not have: cf. Dabrowska 1997) and underestimate their
lexical accomplishments (what is actually involved in learning a word) thus
Words 113
sending several generations of generative linguists on a dubious quest for

Universal Grammar.
I hope to have demonstrated that locative terms, and by implication, lexical
concepts in general, must be learned, and learning them is no easy matter. It
requires some rather sophisticated detective work, including the ability to
project virtual boundaries which are not given in the input. In spite of this,
children begin to use the spatial terms of their language very early, and their
usage is overwhelmingly correct. This suggests that the early and largely error-
free development of grammatical knowledge need not be taken as evidence
that it is innate: if children are intelligent enough to learn lexical concepts, they
might be intelligent enough to learn grammar as well. In fact, considering the
complexity and sheer magnitude of the task (at least partial entries for 60,000
words!), the reader might well conclude that learning syntax is the easy part.
Of course learning grammar is dierent from learning vocabulary: grammar
isnt just about the relationship between form and meaning, but also about how
smaller units are put together to form more complex expressions. In the follow-
ing chapters, I will address some of these dierences, but I will also argue that
the dierences have been overemphasised. We will see that learning grammar
resembles lexical development in many important respects, and conversely, that
learning words entails a fair amount of syntactic learning.
Notes
1. For example, Miller and Gildea (1987) found that children participating
in vocabulary-building programmes at school produced sentences such as
I was meticulous about falling o the cli and Our family erodes a lot. They
were puzzled by these until they consulted the dictionary that the children
were using and learned that meticulous means very careful and erode, to
eat out.
2. Presumably, one can use both forms when describing a currently invisible
dog: in this case, the speaker would assume the vantage point of a virtual
observer. Virtual observers also gure in the semantics of English expres-
sions. Imagine you are describing a family photograph depicting Grandpa
Watkin standing beside Granny Watkin. You can say either Grandpa
Watkin is to the left of Granny Watkin or Grandpa is on Grannys right,
depending on whether you are assuming the point of view of the person
looking at the photograph or of the people in it.
3. This is not an isolated example of such projection: other hilly topo-
graphic adverbs also have extended senses relating to body parts (Casad,
personal communication).
4. The following discussion is an adaptation and extension of a similar dis-
cussion in Bowerman (1993, 1996). I would like to acknowledge my debt
to her.
5. Historically, Mixtec locatives seem to have arisen through metaphorical

projection of body-part terms onto inanimate objects. It is unclear,
however, to what extent the metaphor is still alive in speakers minds
today. Brugman maintains that speakers do indeed project body-part
structure onto inanimate objects; others (e.g. Levinson 1994; Regier 1996)
have argued that the relationship between the body-part and locative
senses is purely diachronic, and that speakers choose the appropriate loca-
tive term by performing a complex analysis of the shape of the ground
object. However, Levinson studied a dierent language Tzeltal whose
body-part locatives are subtly dierent from those of Mixtec. Unlike
Mixtec, the Tzeltal system is object internal, not relational: that it to say,
the body-part construction species that the gure is located at a partic-
ular part of the ground, regardless of the grounds orientation (for
example, at the ear of the table, i.e. its corner), not that it is located in a
particular position relative to the ground located in a larger reference
frame (e.g. on top of/underneath the table). Thus, at the butt of the bottle
always means at the end opposite the neck, even when the bottle is upside
down or horizontal. Furthermore, when Tzetzal body-part terms are
applied to inanimate objects, they refer to object parts, and cannot be pro-
jected to regions around them: at the nose of means contiguous to a
pointed part of, not in front of. (To talk about regions, speakers must
use other systems: relational nouns or a geocentric system see below.)
Regiers argument is based on the fact that his connectionist network
learned Mixtec locatives in spite of the fact that it knew nothing about
body parts. This shows that it is possible to use body-part locatives appro-
priately without performing the metaphorical mapping; but of course it
does not prove that this is what humans do. Note, too, that the point that
the data was used to illustrate remains valid whatever the synchronic
status of the metaphor: the Mixtec locative system is fundamentally
dierent from the systems employed by familiar European languages.
6. Relative systems, incidentally, also dier in how they project orientations
from objects which do not have an inherent front or back. If an English
speaker stands facing a tree with a ball between her and the tree, she would
say that the ball is in front of the tree: the trees front is the side that is
facing the observer. In some other languages for example, in Hausa
the trees front is the side facing in the same direction as the observer, so
the ball would be behind the tree.
7. Bavin does not tell us what the correct adult response would have been in
each case, so we do not know whether the children had acquired the target
system. It is clear, however, that they were not operating with concepts
corresponding to the English prepositions in, on and under.
8. The word-learning task requires participants to learn a new phonological
form, so it could be argued that it is in fact more dicult than the uncle
task. To see if this aected performance, Markson and Bloom introduced
Words 115
a new non-lexical condition which required participants to store a lexical

form: participants were told that one of the unfamiliar objects came from
a place called Koba. Performance was still good, and after a months
delay, there were no signicant dierences between recall in this condition
and the koba and uncle conditions.
9. The social stereotypes enshrined in the word spinster are no longer appli-
cable, which is why the term is becoming obsolete. On the other hand, we
have a new term which, according to the traditional feature analysis,
should be contradictory: bachelor girl.
10. See Lako (1987) for an in-depth discussion of how our understanding of
the concepts and is bound up with knowledge of complex
domains such as biological parenthood, nurturance, legal responsibilities
etc.
11. This is another aspect of the linking problem discussed in Chapter 6.
8 On rules and regularity
1. Words and rules
As we saw in the last chapter, learning the basic lexical units of a language is
a daunting task. But to know a language, one must also be able to produce and
understand larger expressions: morphologically complex words, phrases and
sentences. Such complex structures can be assembled by combining smaller
units according to the rules of grammar.
Intuitively, grammatical rules and words seem to be very dierent kinds of
things. Words are formmeaning pairings, learned by associating stretches of
sound with meanings. Rules, on the other hand, are procedures for manipu-
lating words or rather, classes of words. A simple, and rather hackneyed,
example is the rule for forming the past tense of regular English verbs, which
adds the -ed sux to the verb stem. Linguists usually represent this procedure
with a formula such as this:
(1) V [Past] V + -ed
Simple though it is, the past-tense rule is abstract, in that it can apply to any
(regular) verb; consequently, unless its application is pre-empted by the exis-
tence of a stored irregular form, speakers can use it to produce the past tense
of any verb in their vocabulary, and of any new verb that they might encoun-
ter.
It is rules which allow us to go beyond the expressions we have actually
heard and produce novel forms. Thus, to become competent speakers of their
language, children must acquire some mental analogue of rules. And in fact,
there is plenty of evidence that they do. Sticking to the English past-tense rule,
a classic study conducted by Jean Berko in 1958, and numerous replications
by other researchers since then, have demonstrated that children are able to
inect novel verbs in experimental conditions: for example, when presented
with a nonsense verb like zib they are able to form the past tense form zibbed.
Outside the laboratory, nearly all children occasionally produce regularisation
errors such as breaked and catched, sometimes after a period of correct pro-
On rules and regularity 117
duction. Until the mid-1980s, nearly everyone agreed that these well-known
facts provided direct evidence that children learn symbolic rules of the kind
postulated by linguists.
2. The connectionist bombshell
This seemingly unshakeable view was challenged in 1986 by Rumelhart and

McClelland, who reported that they had trained an articial neural network
to produce the past-tense forms of English verbs. Their network took phono-
logical representations of verb stems as input and produced phonological rep-
resentations of past-tense forms as output. It had been trained on a large set
of stempast pairs, and was subsequently tested on a new set of verbs which
it had not encountered before.
The network consisted of a set of input units and an identical set of output
units, with each input unit connected to each output unit. The entire assem-
bly could be thought of as an extremely simple model of the brain in which
the units correspond to neurons and the connections to axons and dendrites.
Its most interesting property was that it could learn: the connections varied in
strength and could be modied during training.
Rumelhart and McClelland trained the network by presenting it with a
pattern of activation representing an English verb stem and allowing it to
produce an output that is to say, a pattern of activation on the output units
representing the past-tense form of that stem. The training algorithm then com-
pared the actual pattern of activation of the output units to the target pattern
and modied the connection weights so as to minimise the dierence. This pro-
cedure was repeated many times with all the verbs in the training set until the
network was able to supply the correct past-tense forms of most of them.
In the next phase, Rumelhart and McClelland tested the model using the
verbs in the original training set as well as a new set of verbs which it had not
encountered before. They found that it was able to provide the past-tense form
of many (though not all) of the novel verbs in other words, that it was able
to generalise beyond the verbs it had been trained on. Furthermore, the
network sometimes regularised the irregular verbs it had learned earlier, pro-
ducing forms like catched and digged, even though it had supplied the correct
irregular forms before. Thus, in two important respects, the networks
behaviour resembled that of human language learners.
As noted earlier, the occasional occurrence of regularisation errors in chil-
drens spontaneous speech and the fact that they are able to inect novel words
in an experimental setting have long been considered as direct evidence that
children learn symbolic rules of the kind that have been proposed by linguists
(e.g. add -ed to any V). Such rules make use of abstract symbols, such as noun
and verb, and formal operations, such as suxation, which are specic to
language. The apparent ease with which children learn them has been taken as
evidence that they are equipped with a powerful learning mechanism dedi-
cated specically to acquiring language. Rumelhart and McClellands network
used a relatively simple, general-purpose learning mechanism to extract pat-
terns from the input, and produced the same kind of behaviour simply by
forming low-level associations between representations, without any explicit
rules. This raised the possibility attractive for some, unpalatable for others
that the rule-like behaviour seen in humans might also be a product of such a
general-purpose, low-tech mechanism.
3. The linguists strike back
As critics were quick to point out, there were many problems with this early
model (Lachter and Bever 1988; Pinker 1991; Pinker and Prince 1988). The
network could learn rules that are not attested in any human language (e.g.
form the past tense by mirror-reversing the order of segments), but it could
not learn some fairly common patterns (such as reduplication of the stem). It
sometimes produced blends which sound extremely strange to human ears
(e.g. membled for mailed, conafted for conicted). Because it only represented
phonological properties, it could not handle homophones with dierent past
tense forms such as lie to recline (past tense lay) and lie tell lies (past tense
lied). It was strongly dependent on phonological similarity, and hence unable
to inect novel regulars which did not resemble regulars in the training set.
Most damning of all, critics argued, Rumelhart and McClelland got their
model to generalise the regular pattern by manipulating the input. The
network was initially trained on just ten verbs, eight of which were irregular.
When these had been learned, Rumelhart and McClelland increased the train-
ing set to 420 verbs, of which the overwhelming majority were regular. It was
this vocabulary discontinuity or, to be more precise, the fact that it was sud-
denly swamped by regular mappings that caused the network to generalise.
Some of these problems were addressed in later models (e.g. Daugherty and
Seidenberg 1992; Hare and Elman 1995; Hare et al. 1995; MacWhinney and
Leinbach 1991; Marchman 1993; Plunkett and Marchman 1991, 1993). The
later models incorporated one or more layers of hidden units between the
input and the output units, allowing the network to form internal representa-
tions, and hence making it less dependent on surface similarity. They also
adopted more realistic training regimes, and some represented aspects of
semantic structure, which enabled the network to distinguish homophones.
However, the critics also struck at an idea that lies at the very heart of the
connectionist approach to modelling morphological rule learning namely,
the claim that regular and irregular inections are handled by a single mech-
anism by drawing attention to signicant psychological and neuropsycho-
logical dierences between regular and irregular forms (see Pinker 1998, 1999;
Pinker and Prince 1992; Prasada and Pinker 1993). As pointed out earlier, chil-
dren learning English sometimes apply the regular inection to irregular

stems, producing errors like comed and bringed. On the other hand, errors of
the opposite kind i.e. irregularisations such as bat for bit (on analogy with
sat) or truck for tricked (on analogy with stuck) are exceedingly rare.
Likewise, when overgeneralisations are induced during psycholinguistic
experiments with adult participants (for example, by forcing them to supply
the past-tense form under severe time pressure), regularisation errors are
vastly more frequent than irregularisations.
Secondly, irregular verbs are more sensitive to frequency than regulars. In
psycholinguistic experiments, reaction times to low-frequency irregulars are
slower than reactions to irregular forms of high frequency. Also, when forced
to produce past-tense forms very quickly, people tend to make more errors
with low-frequency irregulars (Bybee and Slobin 1982). Results for regular
past-tense forms are mixed. In some studies (Pinker and Prince 1992; Ullmann
1999) participants reacted equally fast and equally accurately to high-fre-
quency and low-frequency regulars. Other studies, however, do report fre-
quency eects for regulars, although, in English, these were observed only for
relatively high-frequency forms (Alegre and Gordon 1999; see also the discus-
sion in Chapter 2, section 3.1.4).
Thirdly, regular and irregular inections may be dierentially aected in
various special populations (Marslen-Wilson and Tyler 1997, 1998; Ullman,
Corkin et al. 1997). Anomic aphasics and Alzheimers patients seem to have
particular problems with irregularly inected words, while regular processes
are largely unaected. In contrast, in agrammatic aphasics, Parkinsons
patients, and children with SLI the regular system is disrupted, while perfor-
mance on irregulars remains relatively good.
Finally, some imaging studies (Jaeger et al. 1996; Ullman, Bergida and
OCraven 1997) suggest that dierent areas of the brain are active when pro-
cessing regular and irregular inections. These results, however, are proble-
matic, as it is not always the same areas that are associated with the processing
of regular verbs on the one hand and irregular verbs on the other in dierent
studies; moreover, not all researchers have found such dierences (see
Seidenberg and Hoener 1998).
4. The dual-mechanism model
We have seen that there is good evidence that regular inections are dierent
from irregulars in several important ways although it is worth noting at this
juncture that most of the evidence relates to just one morphological sub-
system, namely, the English past tense. On the other hand, even die-hard
critics of connectionism like Pinker and Marcus acknowledge that articial
neural networks can model some aspects of human linguistic knowledge and
are able to generalise previously learned patterns to novel input. These critics
argue, however, that the productivity exhibited by network models is fairly

limited because it is constrained by phonological similarity: the models gen-
eralise a pattern to a novel word when it resembles something they have
encountered before. This kind of productivity, Marcus and Pinker suggest, is
characteristic of irregular processes in human language. They are adamant
that connectionist networks cannot adequately model regular inections,
which do not depend on memory for previously encountered forms and can
apply to any stem, regardless of its phonological properties. Regular inec-
tions, they propose, rely on a dierent psychological mechanism: symbolic
rules. To explain human behaviour, then, we need two mechanisms: an algo-
rithmic device capable of implementing symbolic rules, which is responsible
for regular inections; and associative memory, which supplies irregular forms
and, occasionally, novel forms which resemble irregulars (Marcus et al. 1995;
Pinker 1998, 1999).
Thus, the basic premise of the dual-mechanism model is that regular inec-
tions are dierent because they can apply as a default to any stem. Children
are able to learn the regular inection because they are endowed with a mental
mechanism specically dedicated to manipulating abstract symbols; connec-
tionist models, which lack such a mechanism, cannot acquire a true default.
But is it true that they cannot? This might have been true of the Rumelhart
and McClelland model, but not of more recent models which incorporate
hidden units. Hidden units, as pointed out earlier, allow networks to form their
own representations of the input, which makes them less dependent on surface
similarity. Furthermore, connectionist nets are very sensitive to frequency
information, and tend to develop stronger representations for patterns which
are frequently represented in the input. Because regular inections apply to
more words than irregular patterns, the nodes representing them become very
excitable, and hence more likely to become active when a novel stem is pre-
sented. Thus, networks can exploit statistical properties of the language to
learn the regular inection.
But this is not good enough for proponents of the dual-mechanism theory,
who insist that regularity in the psychological sense has little to do with fre-
quency. It is true that regular inections tend to apply to more stems than
irregular patterns but this is a consequence, not the cause, of their being
regular. Regular inections become common because they are applied to new
words that enter the language, and occasionally to low-frequency irregulars.
In other words, they are not regular because they are frequent; they are fre-
quent because they are regular.
What then is regularity in the psychological sense, and how do speakers
know that an inection is regular? Since regular inections do not depend on
access to memory, the theory predicts that the regular inection will apply in
all circumstances in which memory cannot be or is not accessed. Marcus et al.
(1995) and Pinker (1998) list a number of such circumstances; they are sum-
marised in Table 8.1.
Table 8.1 Circumstances in which memory is not accessed and the default
inflection is applied (adapted from Marcus et al. 1995)
Circumstance Examples
Novel words wugs, zibbed

Low-frequency words stinted, eked
Borrowings latkes, cappuccinos
Unusual-sounding words ploamphed, krilged
Words with irregular homophones lied (cf. lay), hanged (cf. hung)
Words which rhyme with irregular words blinked (cf. sank), glowed (cf. grew)
Names the Childs, the Manns
Truncations synched, specs
Acronyms PACs, OXes
Onomatopoeic words dinged, beeps
Words which are mentioned rather than I checked the article for sexist writing and
used found three mans in the first paragraph.
Derivation from dierent category:
(a) denominal verbs spitted put on a spit
(b) deadjectival verbs righted returned to an upright position
(c) nominalisations ifs, ands, buts
Derivation via dierent category:
(a) via noun (V N V) costed calculated the costs
(b) via verb (N V N) wolfs instances of wolfing
Derivation via name:
(a) eponyms Mickey Mouses, Batmans
(b) products Renault Elfs
(c) teams Toronto Maple Leafs
Bahuvrihi compounds sabre-tooths, low-lifes
The clearest example of such a situation is when speakers inect nonce

words like wug or zib in a psycholinguistic experiment. Since such words are
made up by the experimenter for the purpose of the experiment, speakers
obviously cannot rely on stored representations of their inected forms; yet
they happily provide regular past-tense forms like zibbed and regular plurals
such as wugs. A similar situation arises in real life when speakers have to inect
low-frequency words or words which have only recently been borrowed from
another language; and these, too, receive regular endings.
The next three items in the table illustrate the fact that the use of the regular
inection does not depend on similarity to stored exemplars. Speakers apply
the regular infection to unusual-sounding words like ploamph or krilg, for
which there are no similar-sounding entries which they could analogise from,
and also to verbs that are similar or even identical in sound to well-established
irregular forms. Thus, the past tense of blink is blinked, not blank, in spite of
the existence of irregular forms like stank, sank and drank; and the past tense
of lie to tell untruths is lied, not lay, in spite of the fact that its homophone
lie to recline is irregular. Hence, Yesterday I lay on the sofa and lied to my boss.
The regular inection is also used with various non-canonical roots, includ-
ing names, truncations, acronyms, and phonological sequences which repre-
sent noises or which stand for the word itself rather than its referent that is
to say, when the word is mentioned rather than used. (The sentence I checked
the article for sexist writing and found three mans in the rst paragraph means
that the speaker found three occurrences of the word man, not three adult
male human beings pressed between pages 1 and 2: the form man refers to the
word, not its referent.)
Finally, the regular inection is used with several categories of derived
words. Typically, derived words inherit the grammatical properties of their
head, so if the latter is irregular, the derived word, too, is irregular. Thus, the
past tense of overdo is overdid because the past tense of its parent, the verb
do, is did; and the plural of stepchild is stepchildren because the plural form of
child is children. Some derived words, however, cannot inherit their grammat-
ical properties from their parent because they have the wrong kind of parent.
For example, a verb derived from a noun or an adjective cannot inherit its
past-tense form from the original word because nouns and adjectives do not
have past-tense forms. When this happens, the derived word is inected regu-
larly: the past-tense form of the verb to spit, meaning to put on a spit, is
spitted (not spat); and the past tense of to right return to an upright position
is righted (not rote or wrote). This principle also applies to nouns derived from
other categories: we speak of ins and outs and ifs and buts.
Some words have more convoluted and more interesting genealogies.
Sometimes a verb is derived from a noun which itself was derived from a verb
(in other words, the verb has a noun as parent and a verb as grandparent), or
a noun might be derived from a verb which was derived from a noun. What is
interesting about such cases is that the intervening category (the middle gen-
eration) makes it impossible for the new word to inherit the inectional prop-
erties of its grandparent. For example, the verb cost calculate the costs, as in
She costed the proposal, is derived from the noun cost the amount of money
needed for a particular purpose, which in turn comes from the irregular verb
cost meaning be worth a particular amount of money. However, cost calcu-
late the costs cannot inherit the irregular past-tense form from cost be worth
a particular amount because it was derived via the noun, and nouns are not
marked for tense; consequently, the information about the irregularity of the
original verb is not accessible to the processing mechanism, and the default
inection must apply. The same phenomenon can be observed in words
derived via a name. The common noun Mickey Mouse (meaning simpleton)
is derived from the name of the Disney character, and since names dont have
plural forms, it cannot inherit the irregular plural of its grandparent, the
common noun mouse. Hence, the plural of Mickey Mouse is Mickey Mouses,
not Mickey Mice.
The last item on the list bahuvrihi compounds is another category of

words with a mixed-up family background. Bahuvrihi compounds are com-
pounds which refer to a dierent class of things from either of the com-
pounded words: for example, a pickpocket is a kind of thief, not a kind of
pocket or of picking; a bluebell is a kind of ower, not a bell, and so on. The
plural forms of bahuvrihi compounds in English are always regular, even
when the rightmost element is irregular: sabre-tooths (not a tooth but a tiger),
low-lifes (not a life but a person).
As the reader has no doubt noticed, the circumstances listed in Table 8.1 are
an odd collection, and some of them are rather exotic. It is also clear that the
use of the default inection in these circumstances does not depend on phono-
logical similarity to stored forms. Therefore, Pinker and Marcus argue, con-
nectionist models, which map phonological representations of stems onto
phonological representations of target forms, will not be able to learn to
supply the regular inection when required unless, of course, they are spe-
cically designed to deal with each of these circumstances individually.
Learners equipped with symbolic rules in addition to associative memory
never face this problem. Once they have identied the regular inection, they
apply it as a default in all circumstances except when they have specically
learned to use an irregular pattern. In other words, they dont need any special
knowledge about the circumstances in Table 8.1, whether innate or acquired:
the whole list falls out naturally from the division of labour between the
symbol-manipulation device and associative memory.
Or does it? It is possible that English speakers almost always use -s and -ed
to signal plurality and past tense in the circumstances described earlier
because they are equipped with a special mental device which takes over in
such circumstances. On the other hand, it is also possible that they apply the
regular inection simply because, in practice, it is the only one available. The
irregular past-tense and plural inections in English are restricted to specic
words or very narrow classes of words having certain phonological properties,
and hence are not readily extendable to new contexts. Thus, the fact that the
default endings are generally used in the special circumstances enumerated
in Table 8.1 could be due simply the lack of any other truly productive pattern
in the language.
One of the consequences of the fact that the irregular past tense and plural
in English are extremely restricted in their applicability is that the regular
inections apply to the vast majority of types. This introduces another con-
found: speakers may be more willing to extend the regular inection to new
words simply because it is more frequent, and hence more deeply entrenched
in their minds, than any of the irregular patterns.1 As indicated earlier, accord-
ing to the dual-mechanism theory regularity in the psychological sense has
little to do with frequency; but the fact is that the vast majority of English
nouns and verbs are regular.
There is yet another confounding factor in addition to the dierences in
frequency and applicability. In both the plural and the past tense in English,
the regular and the irregular inections rely on dierent morphological pro-
cesses. Regular past-tense and plural forms are formed by suxation, which
results in inected forms which are easily segmentable into a morpheme
which signals a type of activity and a morpheme that signals tense (e.g. play
+ ed) or a morpheme that designates a type of thing and a morpheme that
indicates plurality (cat + s). Most irregular inections, in contrast, involve
stem changes: ringrang, catchcaught, manmen, mousemice. The stem-
change mechanism is less transparent, and hence probably more dicult to
generalise. Furthermore, because there are several dierent patterns of stem
changes competing with each other, learners are forced to attend to the indi-
vidual properties of each word, rather than to similarities between words, and
this is clearly not conducive to unconstrained generalisation.
To summarise: it is undeniable that regular inections in English are very
dierent from the irregulars. Unlike the irregular patterns, they are not
restricted to a narrowly dened group of stems, but apply elsewhere, when
no other pattern can be used; and there are sharp contrasts in the way that
regular and irregular inections are acquired, processed and aected in lan-
guage impairment. What is not clear is whether these dierences are a result
of the fact that regular and irregular inections rely on dierent mental mech-
anisms, or whether they are attributable to the idiosyncratic properties of the
English system of marking past tense and plurality specically, to the eect
of dierences in frequency, applicability, morphological process, or some
combination of these factors. Because of these confounds, the English past-
tense and plural inections are simply not very good testing grounds for the-
ories dealing with the mental status of rules.
But the dual-mechanism theory, of course, is not just about English inec-
tions: it is about language in general. The circumstances in which the default
inection applies are supposed to fall out automatically from the organisation
of the human capacity for language; hence, the same set of circumstances
should be associated with the regular inection in all languages. What we must
do, then, is to test this prediction against data from other languages. Ideally,
we would want to nd an inectional system in which there are at least two
truly productive endings and in which the regular inection (i.e. the inection
that speakers use in the circumstances enumerated in Table 8.1) applies to a
minority of types. The existence of such minority default systems would
show conclusively that regularity does not depend on high type frequency, and
hence provide strong evidence in favour of the dual-mechanism theory.
Do we know of any such systems? Marcus et al. (1995) argue at length that
in the German plural and past-participle inections the regular patterns
applies to a minority of forms, and oer the Arabic plural as a third example.
Unfortunately, the last two of these putative cases are problematic for two
reasons. First, it is not clear that the regular patterns in these subsystems do
in fact apply to a minority of types (see Boudelaa and Gaskell 2002; Bybee
1995), and in any case, they certainly apply to more types than any one irreg-
ular pattern, of which there are a number in both languages. Secondly, like the
English past-tense and plural inections, these systems use dierent morpho-
logical mechanisms for regulars and irregulars. The past participle of regular
(or weak) verbs in German is formed by adding a sux (-t) and in most cases
also a prex (ge-), while most irregular (or strong) past participles, like irreg-
ular past-tense and past-participle forms in English, require stem changes
(and a dierent sux, -en, as well). In the case of the Arabic plural, the regular
mechanism (the so-called sound plural) simply adds an ax to a stem, while
the irregular mechanism (the broken plural) involves modifying the stem in
various ways, sometimes quite dramatically. The German plural, on the other
hand, appears to be a perfect test case for the dual-mechanism theory, and it
is to this inection that we now turn.
5. The German plural: a minority default?
German has ve plural suxes (-e, -en, -er, -s and zero), three of which (-e, -er
and zero) are sometimes accompanied by umlauting of the stem. Clahsen
(1999), Clahsen et al. (1992), Marcus et al. (1995), Pinker (1999) and others
have pointed out that one of the suxes, -s, has a special status. Like the
regular plural ax in English, it is used in the special circumstances enumer-
ated in Table 8.1: with novel words, names, borrowings, onomatopoeic words,
acronyms, etc. This, they argue, shows that -s is the regular or default ending
for the German plural.
What makes the -s ending particularly interesting is the fact that it applies
to a small number of nouns only about 4 per cent (Marcus et al. 1995: 227)
which makes it a minority default par excellence. Furthermore, most other
plurals are also formed by suxation, thus eliminating the morphological
process confound; and at least two other endings are productive that is to
say, they apply to fairly large and open-ended classes of nouns.
Proponents of the dual-mechanism theory have proposed that German-
speaking adults and children learning German know that -s is the default
ending and therefore apply it whenever a plural form cannot be retrieved from
memory. They support this claim with three types of evidence. First, children
learning German often overgeneralise -s; and although they also overgeneral-
ise the other endings, regularisation errors (i.e. overgeneralisations of the
regular ending) are an order of magnitude more frequent than irregularisa-
tions (Clahsen et al. 1992: 247). Secondly, adult speakers tend to use -s with
novel words, particularly unusual-sounding novel words (Marcus et al. 1995).
Finally, some researchers have argued that -s plurals are processed in a qual-
itatively dierent way (see Clahsen 1999 and the work cited there).
Hence, the German plural has come to be regarded as the denitive evidence
for the dual-mechanism theory, the exception that proves the rule and also
the nal blow for connectionist models of inection. Connectionist networks

are strongly dependent on frequency and consequently, say the dualists, they
are unable to learn to generalise an infrequent pattern freely. Furthermore,
since they have no default mechanism to fall back on when memory cannot be
accessed, they would be unable to apply the regular ending in the special
circumstances listed in Table 8.1, unless they were specically trained or actu-
ally hard-wired to deal with every single circumstance individually (Marcus et
al. 1995).
It should be pointed out at this juncture that the claim that connectionist
networks are unable to learn a minority default is actually incorrect, as shown
by network simulations described by Hare et al. (1995), Plunkett and Nakisa
(1997) and Pulvermller (1998). However, it is true that systems of this kind
are more dicult for networks to learn; and clearly, networks trained on
phonological representations of stempast or singularplural pairs cannot
know that conversions, words which are mentioned rather than used etc., all
require the default ending, even when homophonous with words that take
irregular inections. We will return to these and other problems with network
models later in this chapter; in this section, I focus on the problems that the
German plural poses for the dual-mechanism theory.
First, the evidence that German speakers treat -s as the default is far from
conclusive. Let us begin with the language acquisition data. Clahsen et al.
(1992) report that in their sample, the overgeneralisation rate for -s (25 per
cent) was somewhat higher than the corresponding gure for -en (22.5 per
cent) and an order of magnitude higher than overgeneralisation rates for -e
and -er (2.73 per cent and 0.95 per cent, respectively). This nding is at odds
with other studies of the acquisition of plural marking in German, which
report that -en and -e overgeneralisations are much more frequent than -s
overgeneralisations, particularly in younger children (see Behrens 2001, sub-
mitted; Ewers 1999; Gawlitzek-Maiwald 1994; Kpcke 1998; Mills 1985; Park
1978).2 Moreover, on closer scrutiny, it emerges that the high gure for -s over-
generalisations was arrived at by using a non-standard method of calculating
overgeneralisation rates. These are usually computed by dividing the number
of overgeneralised forms by the number of opportunities for overgeneralisa-
tion, that is to say, the number of plural forms that require some other ax.
Clahsen et al. calculated what we might call relative overgeneralisation rates:
for each ax, they divided the number of overgeneralised forms with that ax
by the total number of forms (overgeneralisations and correctly inected
stems) with that ax. Thus, the gure they cite does not mean that -s was used
with 25 per cent of the words that required some other ax, but that 25 per
cent of all -s-suxed plurals in their data were overgeneralisations. This shows
that -s overgeneralisations are quite frequent relative to the frequency of this
ax. However, if we use the more common method of calculating overgener-
alisation rates, the gure for -s drops to 1.9 per cent, which is rather similar to
the overgeneralisation rate for -e (1.2 per cent) and much lower than that for
-en (11.7 per cent). In other words, -s overgeneralisations are more frequent
than we would predict on the basis of the frequency of the ax alone, showing
that frequency is not the only factor determining overgeneralisation rates; they
are not, however, more frequent in any absolute sense.
In fact, overgeneralisation rates for -en, the most frequent ax, are very high
whichever method we use to calculate them: in Clahsens own corpus, the rela-
tive overgeneralisation rate for -en is 22.5 per cent almost the same as that for
-s. Clahsen et al. do not regard this as a problem; in fact, they manage to turn
it into an argument in favour of the dual-mechanism theory, arguing that chil-
dren overgeneralise -en because they think it is the default ending. As evidence
for this, apart from the high overgeneralisation rates, they point out that
regular plurals do not occur inside compounds,3 and the children in their
sample who had high overgeneralisation rates for -en tended to leave it out in
compounds, even when the adult form contained a plural marker. For example,
instead of saying Dose-n-ner can-s-opener, Bauer-n-hof farmer-s-yard,
Kche-n-fenster kitchen-s-window, their subjects said Dose-ner, Bauer-hof
and Kche-fenster. However, there are some serious problems with this argu-
ment. First, the children did not invariably leave -en out in compounds: in fact,
they supplied it most of the time. Secondly, the -en in compounds does not nec-
essarily have a plural meaning: note, for example, that Bauernhof is prototypi-
cally a yard belonging to a single farmer, not several farmers; likewise,
Kchenfenster is a window in a kitchen, not kitchens. This suggests that the -en
in compounds is not a plural marker, but either a historical residue or a linking
morph inserted for phonological reasons (cf. Wegener 1999). Last but not least,
the nding that children are particularly prone to overgeneralise the plural
marker which they leave out in compounds has not been conrmed by other
researchers (see Gawlitzek-Maiwald 1994).
The claim that German adults treat -s as the default is equally problematic.
It is true that speakers tend to use -s more often with nonce words than with
real words. However, they do not apply it indiscriminately to all, or even most,
novel forms. In Kpckes (1988) experiment, for example, -s plurals accounted
for only 17 per cent of the responses. Furthermore, speakers used dierent
axes with dierent nonce words, depending on the words gender and
phonological make-up: for example, speakers preferred -en with feminine
nouns ending in -ung and -schaft and masculine nouns ending in schwa, -e
with masculine nouns ending in -ling, -s with nouns ending in a full vowel, and
so on. By and large, the likelihood of a particular plural marker appearing
after a particular nonce word depended on how frequently that marker
occurred with real words similar to the nonce word (i.e. real words with the
same endings, the same number of syllables, and so on). Other researchers (e.g.
Gawlitzek-Maiwald 1994) have obtained similar results. Thus, nonce-word
experiments show that German speakers are aware that each of the ve plural
axes has a dierent domain of application, and that they apply the axes
accordingly.
Another source of evidence for the special status of -s comes from two
psycholinguistic experiments tapping speakers performance online, i.e. while
they are actually processing plural forms. The rst of these, by Clahsen et al.
(1997), found that people processed high-frequency irregular nouns faster than
low-frequency irregulars while no such dierences were found with regular
nouns. The second, by Sonnenstuhl et al. (1999), used the cross-modal priming
technique to investigate the eect of singular versus plural primes on lexical
decision tasks involving singular nouns. In experiments using this technique,
people are asked to decide as quickly as possible whether a visually presented
target (in this case, a singular noun) is a word. It is well known that reaction
times are shorter when participants hear a related word (called a prime) imme-
diately before seeing the target, and the experimenter can vary the kind of
prime and measure the eect on reaction times to the target word. Sonnenstuhl
et al. used two kinds of prime: the noun stem itself and the plural form of the
noun. Their results showed that a plural form with -s was as good a prime for
the singular as the singular form itself (reaction times to a singular form like
Karton box were the same when it was preceded by Karton or by Kartons); but
for -er plurals, the singular was a better prime than the plural (participants
reacted faster when Kind child was preceded by Kind than when it was pre-
ceded by its plural Kinder). The authors interpretation of this result is that reg-
ulars, but not irregulars, are morphologically decomposed during processing:
in other words, when processing regular nouns, speakers obligatorily access the
stem form and hence react as if they had actually heard the stem.
Several researchers have pointed out various problems with these experi-
ments, some methodological and some pertaining to the interpretation of the
results (Drews 1999; Sereno et al. 1999). Perhaps the most serious problem is
that both studies compared -s with -er, the one plural ax which everyone
agrees is not productive. Of all the German plural axes, -er is the most
restricted in applicability (it is used only with strong masculine and neuter
nouns) and the least predictable in distribution, so each -er plural must be
stored in the lexicon. It is the only ax which was almost never used by sub-
jects in Kpckes (1988) nonce-word production task, and in an acceptability
judgement study conducted by Marcus et al. (1995), most nonce words with
-er were judged to be the least acceptable of all axed plurals. Finally, -er is
very rarely overgeneralised by children (Behrens submitted; Ewers 1999;
Gawlitzek-Maiwald 1994; Kpcke 1998). Thus, the dierences between -s and
-er could be due to the special characteristics of -er rather than the properties
of the default plural. To test the predictions of the dual-mechanism theory, it
would have been much more appropriate to compare -s with the more produc-
tive axes (-en and -e).
Yet another problem with treating the German -s as a default analogous to
the English -s plural is that, unlike its English counterpart, the German ending
does not invariably apply in all the circumstances in Table 8.1. One systematic
exception, acknowledged by Marcus et al. (1995), is bahuvrihi compounds,
which take whichever plural ending is required by the rightmost element: for
example, the plural form of Gromaul big mouth, braggart is not *Gromauls
but Gromuler because the plural of Maul mouth is Muler. Another
obvious exception is rare words: since -s applies to only 4 per cent of German
nouns, the vast majority of all nouns, rare or frequent, take other plural axes.
In the remaining circumstances, -s is the usual ending, but it is not the only
option, a fact which raises serious doubts about the viability of the dual-mech-
anism explanation. Consider plural marking on borrowings, for example.
Marcus et al. (1995), citing Kpckes (1988) data on recent borrowings, point
out that about 50 per cent of them take the -s plural. They regard this as satis-
factory evidence that -s is used with borrowings. However, as a matter of arith-
metical necessity, if 50 per cent of borrowings take -s, the other 50 per cent do
not. With such a high proportion of exceptions, plural marking on borrow-
ings should be something for Marcus and his co-authors to worry about, not
cite as support for their theory. Note, too, that a large proportion of recent
borrowings in German are either English words or words that entered German
via English. Since many German speakers, particularly of the younger gener-
ation, have a good working knowledge of English, it is not unreasonable to
assume that at least some of the new words were borrowed with their English
plural. Languages do sometimes borrow foreign plurals along with the singu-
lar (cf. English criteria, alumni, seraphim), and this is more likely when the
sux also exists in the borrowing language.
Other exceptions include nominalised verbs, which inect like ordinary
nouns (Essen- meals, from essen to eat; Schau-en shows, from schauen to
look; Kuf-e purchases, from kaufen to buy), and some nominalised VPs
(e.g. Tunichtgut-e neer-do-wells, Habenichts-e have-nots). It is not true,
therefore, that nouns derived from other categories always take the regular
plural. Nor is -s invariably used with names and abbreviations. The latter can
take -en when the nal vowel is stressed: for example, the plural of LPG, from
Landwirtschaftliche Produktionsgenossenschaft farming cooperative, is either
LPGs or LPGen (Wegener 1994: 269). And although -s is strongly favoured with
names, the preference is much stronger with surnames than with place names
(which often take - rather than -s), product names (which can take -, and
sometimes also -en or -e: see Wegener 1999), or rst names (so we have Emil-e
and Beate-n alongside Emil-s and Beate-s). Finally, -en and -e plurals are
applied in default circumstances when the noun stem ends in -s (Indefrey 1999).
To conclude: it is undeniable that -s has a somewhat special status among the
plural axes in German. The acquisition and nonce-word data shows that it is
generalised more readily than one would expect, considering its low frequency,
and it is the usual plural ending in most of the circumstances listed in Table 8.1.
Both of these ndings call for an explanation.4 However, the fact that other
endings are also found in default contexts raises serious doubts about the expla-
nation put forward by proponents of the dual-mechanism theory; and there is
little evidence that German speakers, either child or adult, treat it as a default.
The German plural, then, is certainly not the denitive proof of the dual-
mechanism theory that it is sometimes claimed to be. To evaluate the claims
made by the theory, we need to examine still more data. In the following sec-
tions, we will consider two further inectional subsystems the Polish geni-
tive and dative whose properties, as we shall see, make them particularly
interesting test cases.
6. The Polish genitive: an inectional system without a default
6.1 The system

The Polish genitive marking system is complex and fairly irregular. The geni-
tive singular is normally signalled by one of three endings, -i (or its variant
-y),5 -a or -u. Some nouns, however, take adjectival endings (-ego for mascu-
lines and neuters, -ej for feminines), and some do not decline at all (see below).
Which ending is actually used depends on several factors. The single most
important of these is gender, which can be fairly reliably predicted from the
nominative singular form.6 Feminine nouns almost always take -i or -y
(depending on the phonological properties of the nal consonant of the stem),
neuter nouns usually take -a, and masculine nouns usually take either -a or -u.
All three endings are readily applied to new words, but each is restricted to
a particular subset of nouns. Thus, there is no single sux which could be con-
sidered the regular or default genitive ending for all nouns. However, the fem-
inine and neuter inections are regular in the sense that they apply to the vast
majority of feminine and neuter nouns respectively, and this generalisation
can be easily captured by means of two symbolic rules, one which adds -i/-y
to any stem with the features [N, GENITIVE, FEMININE] and another
which adds -a to any stem with the features [N, GENITIVE, NEUTER].
On the other hand, the distribution of the two masculine endings, -a and -u,
is fairly arbitrary, although there are some broad regularities. Almost all
animate nouns, as well as most nouns which designate tools, body parts, the
names of the months, units of measurement and most brand names require
-a, while nouns designating substances, large objects, locations and abstract
concepts generally take -u. Some derivational axes and stem-nal conso-
nants and consonant clusters are also associated with one or the other of the
two endings (see Kottum 1981; Westfal 1956).7
Most of these criteria, however, are not very reliable: there are many excep-
tions, and they are often in conict. The noun tyton tobacco designates a sub-
stance, and hence should take -u; but it ends in an a-friendly consonant. In
this case, the semantic factor gets the upper hand: the genitive form is tytoniu,
not tytonia. In contrast, mzg brain designates a part of the body and hence
should take -a, but the nal consonant cluster is strongly associated with -u,
and in this case, phonology wins. The noun rozpuszczalnik solvent provides
an example of a case where morphological and semantic criteria are in con-

ict: the sux -nik favours -a, and the noun designates a substance, so accord-
ing to the semantic criterion, it should take -u. In this case, the morphological
factor is decisive: the genitive form is rozpuszczalnika.
There are some complications in addition to the problems with inanimate
masculines. Some nouns require endings which are normally used with adjec-
tives. These include deadjectival nouns such as uczony scholar (lit. learned),
many native surnames, and foreign names ending in -e (e.g. Goethe, Rilke).
Furthermore, some masculine nouns designating humans decline like femi-
nine nouns, and some nouns, mostly borrowings and foreign proper names,
especially those that that do not resemble Polish words, are not inected at all.
In summary, there is no single default inection which, unless blocked,
applies freely to all nouns, although one could say that feminine and neuter
nouns do have default endings, namely, -i/-y and -a respectively. However, it is
not clear which, if any, of the masculine endings is regular. Traditionally, this
status is assigned to -a, because it is more frequent than any of the other
endings, but according to the dual-mechanism theory, regularity in the
psychological sense (i.e. default status) does not depend on frequency, and
hence we need to consider other criteria.
As explained earlier, one of the central tenets of the dual-mechanism theory
is that regular inections must apply in the default circumstances discussed
earlier, so we can use this set of circumstances as diagnostic contexts for iden-
tifying the default. If only one masculine ending occurs in these circum-
stances, we can conclude that it is the regular ending in the technical sense of
the theory; otherwise, both endings must be regarded as irregular.
6.2 Genitive masculine forms in default contexts

We must begin by noting that some of the criteria listed in Table 8.1 are either
not applicable to the Polish data or do not allow us to distinguish regular and
irregular forms. The former include onomatopoeic words (tik-tak tick-tock,
dzyn-dzyn ding-dong, kukuryku cock-a-doodle-doo etc.) which are neuter
and hence cannot take a masculine ending,8 and names and words derived
from them, which inect for case just like ordinary nouns. The latter category
includes homophones and rhymes. It is relatively easy to nd pairs of lexemes
which share the same phonological form but take dierent endings (2) and
pairs of rhyming words which take dierent endings (3):
(2) a. tenor tenor (voice), gen. tenoru tenor tenor (singer), gen. tenora
b. tl-ok crowd, gen. tl-oku tl-ok piston, gen. tl-oka
(3) a. motor motorcycle, gen. motoru traktor tractor, gen. traktora
b. smak taste, gen. smaku ssak mammal, gen. ssaka
The problem is that the mere existence of such pairs does not help us to deter-
mine which ending is regular, since both homophony and rhyme are symmet-
rical relations: for every -a noun which has a homophone that takes -u, there
is an -u noun which has a homophone that takes -a; likewise, for every -a noun
that rhymes with an -u noun there is an -u noun that rhymes with an -a noun.
The remaining nine criteria, however, can be applied to the Polish genitive, and
it is to these that we now turn. (Novel words and overgeneralisations in child
language will be dealt with in the following subsections.)
Low-frequency words are equally likely to take either ending. Of the 100 least
frequent masculine nouns from a frequency list compiled by Kurcz et al.
(1990), 48 take -a and 46 take -u.
Borrowings usually take -a if animate and mostly -u if inanimate. Some,
however, are not inected at all, usually because they are in some way non-
canonical and hence cannot be assimilated to any of the established templates.
For example, guru and urdu do not sound like Polish nouns, which never end
in -u in the nominative. Kamikadze, attach, and dingo are phonologically
similar to Polish neuters, but they are assigned the masculine gender because
of their meaning (kamikadze and attach prototypically refer to males, and
dingo inherits its gender from its superordinate pies dog). Because of this, the
neuter pattern is not applicable and such nouns remain uninected. The word
boa is even more interesting. It inherits its gender from its superordinate waz
snake, which is masculine. Although words ending in -a in the nominative are
usually feminine, there is a sizeable group of masculine nouns with this ending
(e.g. poeta poet, artysta artist, logopeda speech therapist); these, in spite of
their gender, are inected like feminines. However, such nouns invariably refer
to human beings, and because of this semantic discrepancy, the pattern cannot
be applied to boa, an animal.
The fact that some words, particularly non-canonical-sounding words, are
left uninected is dicult to accommodate in the dual-mechanism theory,
according to which the default inection must apply if there are no similar
examples stored in memory which the speaker could analogise from. One
could argue, of course, that the default ending in this case is zero. The problem
with this is that such non-canonical words are found in all three genders, and
they are left uninected in all cases, not just in the genitive. Following the same
logic, we would have to conclude that zero is the default ending for all cases
and genders. This would amount to saying that in a richly inected language
like Polish, all the real work is done by associative memory, while the rule
system is on hand to add zero that is to say, do nothing when memory fails.
This is clearly not a very attractive proposal.
Unusual-sounding words: as explained earlier, words which cannot be
accommodated in any established template are simply left uninected.
Words which are mentioned rather than used can occur in the citation form
(4), in the form in which they appeared in the text quoted (5), or in the form
required by the syntactic context of the sentence in which they are mentioned
(6). In the latter case, the usual ending is used: -a nouns such as autor author
take -a and -u nouns like rasizm racism take -u.
(4) Jak sie pisze autor?
How do you spell autor (author:NOM)?
(5) Skresl wszystko od autorowi do rasizm.
Cross out everything from autorowi (author:DAT) to
rasizm (racism:NOM).
(6) Nie moge znalezc tego drugiego autora/rasizmu.
I cant nd the other (occurrence of the word) autor
(author:GEN)/rasizm(racism:GEN).
Truncations usually take the same endings as the words from which they are
derived either -a or -u, depending on the base noun. Thus, dyr (from dyrek-
tor director) and merc (from mercedes) both take -a, like their full forms;
while hasz, like its base haszysz hashish, takes -u. When the truncated stem
is not a noun, the choice of ending is determined by semantic factors. For
example, sam supermarket is derived from (sklep) samoobsl-ugowy self-
service (shop). Samoobsl-ugowy is an adjective and hence takes the adjectival
ending -ego in the genitive, which sam cannot inherit because it is not an adjec-
tive and does not sound like one, so in this case, the usual semantic criteria
apply: sam, like most nouns designating large objects most naturally concep-
tualised as locations, takes -u.
Acronyms are often left uninected; but when they do take an ending, it is
nearly always -u. By this criterion, then, the default ending is -u. However, the
overwhelming majority of acronyms designate organisations and institutions,
i.e. abstract entities, so the choice of ending can be explained by appealing to
semantic criteria.
Words derived from a dierent category: when the noun is derived from a
dierent category (e.g. a verb or an adjective), the choice of ending depends
primarily on the morphological process used in the derivation. Most mascu-
line nouns derived from verbs designate either the agent of the action (e.g.
pisarz writer) or the instrument used to perform the action (zszywacz
stapler), while those derived from adjectives tend to designate individuals
possessing a particular characteristic (e.g. dziwak an eccentric from dziwny
strange, eccentric). These nouns invariably take -a. As mentioned earlier (see
note 7), in such cases semantic and morphological factors work together: der-
ivational suxes come to be associated with a particular inectional ending
because of the meaning of the words formed with the sux; and the same
inectional ending is used even when the derived word doesnt have the right
semantic properties. On the other hand, nouns derived using the three
u-friendly axes, -unek, -ot and -izm, take -u. This is also compatible with their
semantics, as nearly all nouns with these axes are abstract.
Nouns can also be derived from words belonging to other categories by
removing the verbal or adjectival sux from the base. Most words derived in
this way take -u: bieg a run from biegac to run, wykl-ad lecture from
wykl-adac to lecture, braz the colour brown from brazowy brown (adj.) etc.
Again, this is largely attributable to semantics, since such words tend to desig-
nate abstract concepts. However, the use of -u with back-formations is not uni-
versal: -a is used when the resulting noun refers to an animate being (e.g. psuj
a clumsy person who is always breaking things from psuc to break, brutal a
brutal person from the adjective brutalny brutal) and very occasionally with
inanimates as well, especially when the derived noun refers to a concrete object
(gryz a bite from gryzc to bite, skret roll-up from skrecac to twist).
The last possibility for deriving new nouns from other categories involves
conversion. This option is available only with adjectives such as uczony
learned or sl-uzacy serving which can also be used as nouns meaning
scholar and servant. Such nouns invariably keep their adjectival inection.
Words derived via a dierent category can take both endings, and the choice
depends on the morphological processes used in the derivation. If the derived
noun is formed with an a-friendly ax like -acz, -nik, or -arz, it takes the ending
preferred by the ax. For example, the noun utleniacz oxidant is derived from
the verb utleniac to oxidise, which in turn was derived from the noun tlen
oxygen, and it takes -a in spite of the fact that tlen takes -u. On the other hand,
nouns derived from denominal verbs by back-formation require -u. Thus,
dodruku is the genitive form of the noun dodruk an additional print run,
derived from the verb dodrukowac to print additional copies, which in turn
comes from drukowac to print, which is derived from the noun druk print.
Bahuvrihi compounds referring to inanimate objects typically take -u,
although there are many exceptions; animates, on the other hand, always take
-a. Here are some examples:
(7) -u inanimates:
piorunochron lightning rod (lit. lightning-protect), gen.
piorunochronu rwnolegl-obok parallelogram (lit. parallel-side),
gen. rwnolegl-oboku
(8) -a inanimates:
katomierz protractor (lit. angle-measure), gen. katomierza
trjkat triangle (lit. three-angle), gen. trjkata
(9) -a animates:
garkotl-uk a clumsy cook (lit. pot-break), gen. garkotl-uka
stawong arthropod (lit. joint-leg), gen. stawonoga
Particularly relevant for our purposes are compounds which have a masculine
noun as the rightmost element but require a dierent ending from the noun.
As the following examples illustrate, neither ending has a special status: a com-
pound containing an -a noun may require -u and vice versa.
(10) -a noun in -u compound:

zab tooth, gen. zeba
trjzab trident (lit. three-tooth), gen. trjzebu
(11) -u noun in -a compound:
strajk strike, gen. strajku
l-amistrajk scab (lit. break-strike), gen. l-amistrajka
The distribution of genitive masculine endings in the various circumstances

which, according to the dual-mechanism theory, call for the default inection is
summarised in Table 8.2. As the summary demonstrates, no single ending is asso-
ciated with all of these circumstances. In fact, most default contexts allow more
than one ending, and which one is actually chosen depends on the properties of
the noun. Since, by denition, the use of the default cannot depend on lexical
properties of the noun, we must conclude that all masculine endings are irregular.
6.3 Adult productivity with genitive endings

It will be clear from the preceding discussion that the confounds which make
the English data dicult to interpret are absent in the Polish genitive marking
system: the irregular inections apply to large and phonologically unrestricted
noun classes and, like the regulars, involve suxation. Thus, the Polish geni-
tive case oers us an excellent opportunity to test a central tenet of the dual-
mechanism theory: that symbolic rules are the main mechanism underlying
our capacity for producing and understanding novel forms, while associative
memory is only marginally productive. If this claim is correct, speakers should
be more productive with the feminine and neuter inections than with the
irregular masculines.
This prediction was tested in a nonce-word experiment in which Polish-speak-
ing adults were asked to supply the genitive form of nouns of all three genders
(see Dabrowska 2004: study 1).9 As explained earlier, there is a strong tendency
for masculine nouns designating small, easily manipulable objects to take -a and
for nouns designating substances to take -u. In order to determine whether Polish
speakers are sensitive to this distinction, one-third of the nonce words were pre-
sented in a sentential context suggesting they referred to concrete objects, e.g.:
(12) Ale malutki gon! Nigdy nie widzial-em takiego malutkiego ______.
What a tiny gon! I have never seen such a tiny ______.
and one-third in a context inviting participants to conclude that they desig-

nated substances, e.g.:
(13) Na przeziebienie najlepszy jest progonys. Wypij dwie l-yzki ______.

If youve got a cold, progonys is best. Take two tablespoons of
______.
Table 8.2 Genitive inflections on masculine nouns in default circumstances

Circumstance Required ending Examples
Low-frequency words -a or -u szanc-a bulwark

czestokol--u palisade
Borrowings -u, -a or fonem-u, pub-u
drink-a, jeep-a
guru-, boa-, swahili-
Unusual-sounding words guru-, boa-, swahili-
Words which are Form in which the word Nie moge znalezc tego
mentioned rather than was originally used, drugiego autor-a/
used citation form, or whatever rasizm-u. I cant find
ending the noun normally the second (occurrence
takes of) author/racism.
Truncations -a, sometimes -u merc-a Mercedes
spec-a specialist
hasz-u hashish
Acronyms and or -u PCK- (Polski Czerwony
abbreviations Krzyz Polish Red Cross)
PAN-u (Polska Akademia
Nauk Polish Academy of
Sciences)
Derivation from dierent
category:
(a) axation -a, sometimes -u zszywacz-a stapler
rasizm-u racism
(b) back-formation -u, sometimes -a skret-u twist, skret-a
roll-up (both from
skrecac to twist)
(c) nominalised adjectives -ego uczon-ego learned,
scholar
Derivation via dierent -a or -u utleniacz-a oxidant
category dodruk-u additional print
run
Bahuvrihi compounds -a for animates; mostly -u l-amistrajk-a scab
for inanimates (break-strike)
trjzeb-u trident (three-
tooth)
The remaining words were introduced as place names, which show no prefer-
ence for either ending, thus providing a base-line condition.
Of the 24 masculine items, four ended in consonants or consonant clusters
which, in the real lexicon, are strongly associated with -a, and four in conso-
nants or consonant clusters strongly associated with -u. The remaining words
had a variety of osets typical for masculine nouns.
Table 8.3 Response types in the nonce-word production study (%)

Expected Other
Gender responses (SD) Gender errors NOM errors responses
Masculine 93.5 (8.4) 2.6 3.9 0.0

Feminine 92.4 (10.7) 1.5 5.0 1.1
Neuter 65.0 (21.8) 16.1 18.3 0.6
Performance on words of each gender (i.e. the number of expected

responses, standard deviations, and the major error types) is summarised in
Table 8.3. As we can see, the number of expected responses for masculine
nouns is no lower in fact, slightly higher than the corresponding gure for
feminines. Performance on neuter nouns, on the other hand, was considerably
worse than on nouns belonging to the other genders. The unexpectedly high
error rates for neuter words were due partly to overgeneralisation of mascu-
line and feminine endings, and partly to use of the citation form (i.e. the nom-
inative). A possible explanation for the poor performance on neuters will be
given in section 7 below.
Out of the 60 participants 57 used both masculine endings productively,
and all nouns were sometimes inected with -a and sometimes with -u, con-
rming that both endings are phonologically unrestricted. However, partici-
pants preferred -a with words which ended in a-friendly consonants, and -u
with words ending in consonants or consonant clusters which are strongly
associated with this ending in the real lexicon: the proportion of -a responses
was 64 per cent with the former and only 25 per cent with the latter. The
respondents showed an even stronger sensitivity to the meaning of the noun
as a determinant of the choice of ending. As shown in Table 8.4, they tended
Table 8.4 Genitive endings on masculine nouns in the Object, Place and
Substance conditions (%)
Ending Object Place Substance Mean
-a 77.3 43.8 19.8 46.9

-u 21.9 41.9 76.0 46.6
Other 0.8 14.4 4.2 6.5
to use -a when the noun was presented as an object and -u when it was pre-
sented as a substance. This pattern was found in 57 respondents; the remain-
ing three gave equal numbers of -a responses in both conditions. Nouns
presented as place names took both endings with equal likelihood, and thus
pattern between the two extremes.
It is clear, then, that regularity in the technical sense of the dual-mechanism
theory is not a good predictor of productivity. Polish speakers freely general-
ise the irregular masculine inections to novel nouns but do not consistently
supply the neuter ending, in spite of the fact that it is regular. Furthermore,
speakers are sensitive to both the semantic and the phonological factors aect-
ing the distribution of masculine endings, conrming that the latter are irreg-
ular in the technical sense of the dual-mechanism theory.
6.4 Acquisition of the genitive inection

Thus, a fundamental tenet of the dual-mechanism theory, the claim that
regular inections are always more productive than irregular ones, turns out
to be incorrect. This, however, does not necessarily entail that speakers rely on
the same mental mechanism to supply both types of inections: it is possible
that two dierent but equally productive mechanisms are involved. If this were
the case, we would expect to nd signicant dierences in the way the two
systems are acquired. It is unlikely that two completely independent systems
mature at the same time, so we should be able to observe developmental asyn-
chronies between regular and irregular inections in other words, children
should be likely to acquire one system earlier than the other. Furthermore, the
rule system and the memory system should show dierent patterns of devel-
opment: specically, we should see a sudden rise in marking rates coinciding
with the onset of overgeneralisation errors for regular inections, and a slow
but steady improvement in performance on irregulars. Finally, we would
expect to nd dierences in overgeneralisation rates similar to those observed
in English-speaking children: children should overgeneralise regular inec-
tions considerably more often than irregular ones. In this section, we will
examine the developmental evidence in order to determine whether any of
these predictions are borne out.
Dabrowska and Szczerbinski (in preparation) used the nonce-word task
with Polish-speaking children aged from 2;3 to 4;9 and found that all genitive
endings are acquired early: even in the youngest group (mean age 2;6), virtu-
ally all children were able to use masculine and feminine endings productively
(that is to say, apply them to at least one nonce word), and about 72 per cent
were productive with the genitive neuter. Moreover, children in all age groups
supplied the correct response more frequently with nonce masculines: in fact,
as in adults, performance on masculine nouns was slightly better than on fem-
inines (although the dierence was not statistically signicant) and consider-
ably better than on neuters (where the dierence was highly signicant);
performance on feminines was also signicantly better than on neuters (see
Figure 8.1).
Thus, the experimental results show no evidence of a developmental advan-
tage for regular inections. This is conrmed by an analysis of longitudinal
spontaneous speech data. Figure 8.2 shows the overall course of acquisition
of the genitive inection in four of the Krakw children, Basia, Inka, Jas and
Kasia.10 The children began to use genitive endings at about 18 months of age.
During the immediately following months, they sometimes used the citation
100
90
80
70
% correct
60
50
40 Masculine
30 Femine
20 Neuter
10
0
2;6 3;6 4;6
Age
Figure 8.1 Childrens performance on the nonce-word production task
form (i.e. the nominative) in contexts requiring the genitive, but such errors
gradually disappeared. As can be seen from Figure 8.2, overgeneralisation
errors were also relatively infrequent, so from about age 2, the childrens use
of the genitive is overwhelmingly correct.
Crucially, the pattern of development of the regular and irregular inec-
tions is very similar. Since the masculine and feminine inections apply to
classes of about the same size (44 per cent and 41 per cent respectively of the
childrens noun vocabulary), the simplest method of measuring growth is to
compare the number of noun types used with a particular inection. Figure
8.3 shows the cumulative number of masculine and feminine nouns used with
a genitive ending during the period from emergence (rst use) until acquisi-
tion (the point at which the child reliably supplies the correct ending) in the
four children. It is clear from the gure that both inections develop at the
same rate and show similar patterns of growth: progress tends to be slower at
the beginning, then speeds up, but neither inection shows a sudden take-o.
Marking rates on masculine and feminine nouns also show similar patterns of
growth (see Dabrowska 2004: study 2).
We now turn to the childrens overgeneralisation errors. According to the
dual-mechanism theory, the vast majority of all errors should be regularisa-
tions: that is to say, children should overgeneralise the feminine -i/-y and pos-
sibly -a (which is the regular ending for neuter nouns as well as an irregular
one for masculines). As shown in Table 8.5, -a was indeed overgeneralised
more often than -u, but this seems to be better explained as a frequency eect:
-a was also 34 times more frequent in the childrens correct productions, and
the relative overgeneralisation rate, which, as explained earlier, factors out the
DABROWSKA text Make up
26/3/04
4:35 pm
Page 140
Figure 8.2 Correct uses and overgeneralizations (OG) of genitive endings in the four children.
Note: The ages on the y-axis represent the endpoint of the relevant period. There is no data for Kasia between the ages of 3;1 and 3;11, so
the horizontal distances on the y-axis are not proportionate to the time between the data points.
26/3/04
4:35 pm
Page 141
Figure 8.3 Cumulative number of masculine and feminine noun types used in the genitive case
Table 8.5 Overgeneralisation errors in the developmental corpora

Overgenerali-
Tokens correct sations AOG ratea ROG rateb
Child -a -u -i/-y -a -u -i/-y -a -u -i/-y -a -u -i/-y
Basia 1,387 1,118 1,422 2 1 0 0.4 0.1 0.0 0.5 0.8 0.0
Inka 1,050 1,424 1,187 9 2 3 0.6 0.1 0.2 0.9 0.5 0.4
Jas 1,137 1,429 1,264 39 10 1 2.3 0.4 0.1 3.3 2.3 0.1
Kasia 1,465 1,119 1,638 17 15 2 2.2 1.3 0.3 3.5 11.2 0.3
Whole corpus 3,039 1,090 3,511 67 28 6 1.4 0.4 0.1 2.2 2.5 0.2
Notes:
a
AOG rate (absolute overgeneralisation rate) = overgeneralisation errors as a
percentage of the number of noun tokens requiring other axes
b
ROG rate (relative overgeneralisation rate) = overgeneralisation errors as a
percentage of the total number of tokens with a given ax
eects of frequency, is very similar for both endings. Crucially, the one ending
which is uncontroversially regular, namely the feminine -i/-y, was overgener-
alised least often.
The children rarely used any endings in gender-inappropriate contexts: that
is to say, they tended to use masculine endings with masculine nouns (though
not necessarily the right masculine ending), the feminine ending with feminine
nouns, and the neuter ending with neuter nouns. As noted earlier, the gender
of most Polish nouns can be reliably predicted from the phonological form of
the nominative, so the infrequent occurrence of gender errors suggests that
children are sensitive to such cues from a very early age. The picture that
emerges, then, is that where there are consistent patterns, children learn them
with ease, and overgeneralisation levels are very low; but where the distribu-
tion of endings is less predictable, error rates are higher. It should be stressed,
however, that these are relative dierences: the overall overgeneralisation rates
are very low, ranging from 0.5 per cent in Basia to 3.3 per cent in Kasia.
Moreover, they remain consistently low throughout the entire period from
emergence until age 5;0 (cf. Figure 8.2).
Such low error rates in a system as complex and irregular as the Polish gen-
itive suggest that childrens performance relies to a considerable extent on the
use of stored forms. A closer analysis of the childrens early usage conrms
this suspicion. As pointed out earlier, there was a period of about ve months
after emergence during which children did not always supply the genitive when
required: some nouns which should carry the genitive inection were left in the
citation form (the nominative). What is interesting is that during this period
of inconsistent usage, genitive and nominative forms were used with dier-
ent populations of lexemes that is to say, at any one point in time, each child
consistently provided the correct genitive form of some nouns and left others
in the citation form. Of course, over time the number of noun types which
were consistently inected in genitive contexts increased, while the number of
noun types which were left uninected got smaller and smaller, so if we con-
sider the entire period, we will nd that some nouns were used in both forms
in genitive contexts. However, in such cases, all the uses of the citation form
generally preceded all the genitive uses. We can see this from Table 8.6, which
summarises case marking on all the nouns in each childs transcripts which
occurred at least twice in genitive contexts (and hence could have been used in
dierent forms) during the period of inconsistent use. In Basias transcripts,
for example, there are 22 noun lexemes which occurred in genitive contexts
during this period at least twice. For eight of the 22 lexemes, there are no
attested uses of the nominative form in genitive contexts: that is to say, as far
Table 8.6 Case marking on noun lexemes used in genitive contexts

Child GEN NOM All NOM At least 1 NOM All types with
only only before 1st GEN after 1st GEN 2+ tokens
Basia 8 7 5 2 22
Inka 32 2 2 1 37
Jas 19 2 4 1 26
Kasia 25 0 8 2 35
as we know, these nouns were used correctly from the very beginning. Another
seven lexemes were never inected during this period: all the attested forms are
nominative. Finally, seven lexemes occurred in both forms. For ve of the
seven, all the recorded nominative forms precede all the recorded instances of
the genitive. For example, there are 21 recorded occurrences of the lexeme tata
dad in contexts which require the genitive. Fifteen of these are in the nomi-
native case, and they all occurred before the genitive form appeared. The
remaining two lexemes were used inconsistently: that is to say, a nominative
form occurred at least once in genitive contexts after Basia had produced a
genitive form of the same noun. There were only six examples of such devel-
opmental U-turns in the entire corpus, and four of these were one-o slips (i.e.
the child used the nominative after a genitive only once). Thus, during the
period when the genitive was not consistently supplied, the childrens perfor-
mance on specic nouns was remarkably consistent, supporting the conjecture
that they were acquiring the genitive inection on an item-by-item basis.
On the other hand, there is also evidence that the children acquired some
rule or schema which allowed them to produce inected forms which they had
not heard before. We have already seen that, from about age 2, they marked
the genitive virtually 100 per cent of the time, which suggests that they were
able to add the inection to all the nouns in their vocabulary. Furthermore,
although overgeneralisation errors were rare, their very existence provides
strong evidence of productive use. The distribution of the childrens errors

also suggests that they were using endings productively: errors occurred most
frequently in the most irregular parts of the system, i.e. on masculine nouns.
However, such generalisations are probably quite specic. For example,
children might learn to use -a with nouns designating animate entities and
things that move, or with nouns having certain phonological properties. More
general rules (e.g. add -a to any masculine noun or add -u to any inanimate
masculine noun) would lead to error, and the fact that errors are rare suggests
that children either do not extract such general rules, or if they do, they nor-
mally rely on more specic rules in production.
7. The nal test case: the Polish dative
The research on the Polish genitive inection reported in the preceding section
brought one unexpected nding: that Polish speakers, adults as well as chil-
dren, are not fully productive with the neuter ending. This is puzzling, since
the genitive neuter inection is one of the simplest and most regular parts of
the case-marking system. Why then do speakers nd this relatively straightfor-
ward inection so troublesome? The answer, it seems, is to be found in the
peculiar structure of the neuter class.
The majority of neuter nouns are morphologically complex, and most
contain one of a few highly productive derivational morphemes. Over a half
of all neuter nouns are action nominalisations formed by adding the sux -nie
to a verb stem; and another 30 per cent are formed with one of three other
highly productive suxes, -cie, -stwo and -ko. As a result, most neuter nouns
are clustered in several small but very crowded regions of phonological space
dened by these axes, and there are relatively few nouns outside these clus-
ters. The masculine and feminine classes, of course, also contain large
numbers of morphologically complex nouns; however, none of the clusters is
so large, relatively speaking; and there are many nouns distributed fairly
evenly throughout the space between the clusters.
If, as suggested above, learners initially extract fairly specic generalisations
over clusters of nouns rather than fully abstract rules, then it is possible that
in the case of neuter nouns, they might never progress beyond these local gen-
eralisations, perhaps because there are not enough nouns outside the clusters
to force them to generalise further. This would result in a situation where
mature speakers would lack a general rule which applies to all neuter nouns,
but would know how to produce the dative forms of words ending in -nie,
-stwo, -ko etc. (or, if they pay attention to meaning as well as form, action
nominalisations formed with the -nie sux, abstract nouns formed with -stwo,
-ko diminutives, and so on). On the other hand, speakers would extract more
general rules in addition to such local regularities for masculine and feminine
nouns, which are more evenly distributed in phonological space.
To investigate this possibility, we will look at another adult nonce-word

experiment but one which tested productivity with dative rather than geni-
tive inections (see Dabrowska 2004: study 3). There are two reasons for
choosing the dative rather than the genitive, or some other inection, to inves-
tigate this issue. First, the dative is the only case in which the neuter ending is
dierent from the case endings characteristic for masculine and feminine
nouns as well as the nominative ending, and it is therefore possible to distin-
guish correct responses, overgeneralisations of masculine and feminine
endings, and failures to apply any inection at all. (As noted in the previous
section, the genitive neuter ending, -a, is the same as one of the masculine
endings, so some of the responses which were coded as correct in the genitive
study might in fact have been uses of the masculine ending.) Secondly, using
the dative inection makes it possible to control for a possible confound,
namely, type frequency. The neuter class is considerably smaller than the mas-
culine and feminine classes: only about 25 per cent of Polish nouns are neuter,
while masculine and feminine nouns make up 41 per cent and 34 per cent of
the noun vocabulary respectively. Thus, neuter endings apply to fewer nouns
than masculine and feminine endings. There is strong evidence suggesting that
type frequency is a major determinant of productivity, so comparing neuter
inections with masculine and feminine inections would produce ambiguous
results: we would have no way of knowing whether any observed dierences
in productivity were due to class structure or simply class size. However, in the
dative case, the feminine class is split between two endings: -i (or its variant
-y), which is used with stems ending in palatalised consonants, aricates and
post-alveolar fricatives, and -e, which is used with all other feminines; and the
number of nouns which take the -e (18 per cent) is roughly equal to the number
of nouns which take the neuter ending.11
All four dative endings (the masculine -owi, the feminine -e and -i/-y, and
the neuter -u) are regular in the technical sense of the dual mechanism theory:
that is to say, all four occur in most default circumstances (see Dabrowska
2004). Note, however, that, contrary to the theory, both feminine endings are
phonologically restricted.
In the experiment, adult native speakers of Polish were given a written ques-
tionnaire testing their ability to supply dative endings with nonce nouns. Of
the nouns, 24 (eight masculines, eight feminines and eight neuters) were
morphologically simple. There were also 16 complex neuter nouns: eight con-
taining the highly productive nominalising sux -nie and eight containing the
diminutive sux -ko, which applies to a much smaller number of nouns. Each
item on the questionnaire consisted of a context sentence which introduced
the nonce word in the nominative and gave a simple denition of the referent,
followed by a test sentence with a blank in a grammatical context which
requires the dative. The items containing morphologically complex words also
contained an additional clause which gave fairly transparent clues about
morphological structure.
The results for the simplex words were similar to those in the genitive study:
participants performed close to ceiling on masculine and feminine nouns, but
gave the expected response to neuter nonce nouns less than 50 per cent of the
time (see Table 8.7). Most errors on simplex neuter nouns involved use of the
masculine or feminine endings (which were overgeneralised with approxi-
mately the same frequency); but in just under 10 per cent of the trials, partic-
ipants failed to inect the nonce word at all, and simply used the nominative
form in a grammatical context requiring the dative. This distribution of errors
was due to all participants responding inconsistently, rather than some partic-
ipants consistently using the same incorrect response with all or most of the
neuter nouns: in other words, the same respondents used the correct neuter
ending with some nouns, the masculine and feminine endings with others, and
left some nouns uninected.
Table 8.7 Response types in the dative study (%)

Gender errors
NOM Other
Condition Expected Masculine Feminine Neuter errors errors
Simplex MASC 89.1 9.0 0.5 1.1 0.3

Simplex FEM 97.6 1.9 0.0 0.0 0.5
Simplex NEUT 46.7 17.9 23.1 9.8 2.4
NEUT-ko 55.7 28.3 14.7 0.3 1.1
NEUT-nie 98.4 0.3 0.0 0.5 0.8
The gures in Table 8.7 also show that participants supplied the expected
ending with morphologically complex neuter nonce words considerably more
often than with simplex neuters; in fact, performance on neuters with the
nominalising sux -nie was even better than on masculines and feminines.
This indicates that speakers do indeed rely on low-level patterns. On the other
hand, participants were able to supply the expected ending with morphologi-
cally simple neuter nonce words in almost half of the trials, which suggests
that they also have a more general rule. This higher-level generalisation,
however, is less well entrenched than the more specic patterns, and too weak
to be reliably used in production.
It is worth noting that performance on -nie nominalisations was much
better than on diminutives. This is almost certainly a type frequency eect: as
explained earlier, there are about ve times as many -nie nominalisations as
there are diminutives with -ko in the real lexicon. Thus, type frequency does
matter, although it does not seem to matter as much as phonological coher-
ence: the class of neuter nouns is necessarily larger than the class of neuter
nouns ending in -nie, yet speakers apparently have a well-entrenched rule for
the latter but not the former. Notice, too, that although the masculine ending
applies to more than twice as many nouns as the feminine -e, performance on
masculines was actually slightly worse than on feminines. This suggests that
the eects of type frequency are non-linear: robust generalisations require a
relatively large number of exemplars; but beyond a certain threshold, further
exemplars do not aect productivity.
8. Interim conclusions
Earlier in this chapter, we saw that research on the English past-tense and
plural marking systems appears to support the dual-mechanism theory. In
both of these systems, a single, phonologically unrestricted ax applies in a
variety of circumstances in which memory cannot be accessed, and there is
convincing psycholinguistic and neurolinguistic evidence that regular and
irregular inections are processed dierently by the healthy brain and dier-
ently aected in neurological disorders. However, it is not clear whether these
dierences reect the properties of two dierent mental mechanisms, or
whether they are simply an artefact of the idiosyncratic properties of the
English system of marking tense and plurality. We also saw that in the German
plural, the distinction between regular and irregular endings is much less clear-
cut. One of the endings, -s, does have a special status, but it is not always used
in default circumstances, and, more importantly, speakers do not treat it as a
default: German-speaking children overgeneralise -en or -e more frequently
than -s, and both children and adults prefer these two irregular endings with
nonce words.
The Polish results cast further doubt on the idea that there is a clear-cut dis-
tinction between regular and irregular inections. We have seen that irregular
inections can appear in default circumstances, and that at least some regular
inections (e.g. the dative feminine endings) are sensitive to phonological
properties. Furthermore, regularity is a poor predictor of productivity: some
irregular inections (e.g. the two genitive masculine endings) are freely gener-
alised to new words, while some regular ones (e.g. the genitive neuter and
dative neuter endings) are not.
The dative experiment also revealed very strong similarity eects on neuter
nouns, in spite of the fact that the dative neuter inection is regular. This is
important in the context of the debate between the connectionists and propo-
nents of the dual-mechanism theory, since it undermines one of the strongest
arguments against connectionist approaches to language. As noted by Pinker
(1999) and Marcus (1998), connectionist models are, by their very nature,
highly reliant on local similarities, and consequently have problems generalis-
ing to novel patterns which do not resemble items they have encountered in
the training set. Research on the Polish case system suggests that this is also a
feature of linguistic generalisation in humans, so this property of connection-
ist models is an argument in their favour, not against them.
Polish speakers restricted productivity with neuter inections also has
important implications for linguistic theory. Most linguists take it as self-

evident that speakers extract the most general rule compatible with the lan-
guage they hear. The results reported here show that this is not necessarily the
case. Although speakers probably have general rules as well, the most robust
linguistic generalisations appear to be local. There is also evidence that people
store the inected forms of many familiar words as well as the citation forms
(see Chapter 2, section 3.1.4; for evidence that Polish speakers store ready-
made dative forms of neuter nouns, see Dabrowskaa 2004: study 4). Thus, our
mental grammars appear to be quite messy, with the same information repre-
sented redundantly at dierent levels of abstraction.
The research on the Polish genitive aords us yet another perspective on
these issues. As shown in section 6.2, neither of the two genitive masculine
endings is regular in the technical sense of the dual-mechanism theory. The
mere existence of an inectional system without a default, though awkward
for the dual-mechanism theory, does not constitute denitive evidence that it
is incorrect. Although the theory implies that inectional systems normally
have a default mechanism, it does not require that every inectional category
in every language have one: it is perfectly possible that some systems in some
languages lack a default. In such cases, speakers would be forced to fall back
on associative memory. There are two problems with this scenario, however.
First, the dual-mechanism theory does not explain how learners discover
that a system does not have a default. Marcus, Pinker and their collaborators
argue that children are programmed to look for the default inection and iden-
tify it either by noting that it is phonologically unrestricted or by observing
that it is used in one or two of the circumstances listed in Table 8.1 (Marcus
et al. 1995: 245; cf. also Clahsen 1999; Marcus et al. 1992: 133). Since both
-a and -u are phonologically unrestricted, and both are found in a number of
default contexts, Polish children could easily conclude that one or the other of
the two endings is the default; and it is unclear how they could recover from
this incorrect conclusion.
Another problem with this defence of the dual-mechanism theory is that it
amounts to saying that children resort to symbolic rules when they are learn-
ing a relatively neat and tidy system such as the English past tense, but have to
rely on associative memory when confronted with a more complex system
such as the Polish genitive. Although such a position is logically consistent, it
is not very attractive, particularly in view of the fact that English-speaking
children learning the past tense appear to make more errors than Polish chil-
dren learning the genitive (Dabrowska 2001). It would seem, then, that the
default system does not do very much, and that English-speaking children
would do better to ignore it and rely simply on memory.
Another important lesson to be drawn from the work on the acquisition of
the genitive is that children are extremely good learners. They master the
Byzantine rules of Polish morphology early and almost without error
(Dabrowska 2001; Smoczynska 1985). They seem to have no great diculty in
restricting each inection to a particular subset of words belonging to a larger

class, even if this subset is essentially arbitrary, at least with very frequent
inections such as the genitive. And they also learn which inection to use in
each of the special circumstances enumerated in Table 8.2. We do not know
when or how they accomplish this, but learn it they must, since dierent
circumstances call for dierent endings.
Marcus and Pinker repeatedly stress that some of the circumstances in
which the default inection applies in English are exotic, and although they
do not explicitly say that they are unlearnable, there is a strong suggestion to
this eect: verbs derived via nouns, bahuvrihi compounds with an irregular
noun in the rightmost position etc., are so rare that children might not encoun-
ter enough examples in the input to learn which inection they require. The
dual-mechanism theory provides a simple explanation of how they come to
know that the default inection applies in each of these circumstances: chil-
dren do not have to learn this, because the whole list falls out automatically
from the architecture of the language-processing system. I do not wish to deny
that English-speaking children do not need to learn which inection to use in
these default circumstances. They whole list does fall out automatically not
from the architecture of the language-processing system, but from the fact
that the regular past-tense and plural endings are the only ones which are truly
productive. Polish children, however, do have to deduce which ending to use
in each circumstance, and the fact that they are able to do this shows that such
contingencies are learnable, however exotic they may be.
9. Challenges for connectionism
The connectionist approach to language and language acquisition has much

to recommend it. To implement their models on a computer, the designers are
obliged to oer fully explicit accounts of the development of a particular
aspect of a speakers linguistic knowledge an account in which all details of
the network architecture, input and output representations, and training
regime have to be spelled out fully. Computer simulations also make it pos-
sible to study the eects of using dierent architectures and dierent training
regimes and to explore the consequences of lesioning the network at dier-
ent stages of development experiments we cannot perform on human beings.
Although connectionist nets are very dierent from real brains, they are
neurologically much more plausible than models inspired by the serial com-
puter, so prevalent in computer science. Most importantly, nets can learn: they
are able to represent recurrent patterns and generalise them to novel situations
and they do this using relatively low-tech, general-purpose mechanisms.
Furthermore, like humans, they are highly sensitive to frequency, extract pat-
terns at varying degrees of abstraction, and are particularly good at learning
local regularities. These similarities suggest that the manner in which networks
generalise by capturing shared properties by means of partially overlapping

representations may be a good model of how humans extract linguistic gen-
eralisations.
In spite of these undeniable advantages, the connectionist approach to lan-
guage has come under severe criticism (see, for example, Lachter and Bever
1988; Marcus 1998; Marcus et al. 1995; Pinker 1999; Pinker and Prince 1988).
While much of the criticism applies to specic models (cf. section 3 above),
other problems are of a more general nature. In this section, I discuss some of
these general problems, possible solutions, and lessons that can be drawn from
them.
One fairly obvious problem with many network models stems from the way
their designers conceptualise the task facing the child. Most modellers assume,
either explicitly or implicitly, that learning to supply the correct inected form
involves learning to relate the phonological representation of the stem and the
phonological representation of the derived form (e.g. the past tense). In line
with this view, networks are trained on stempast pairs like drinkdrank,
sleepslept, likeliked. This assumption, of course, is extremely implausible
when taken at face value: the input available to children obviously does not
consist of stems followed by the correct past tense form. It is possible that on
hearing the past tense form of a verb they already know, children retrieve the
stem from memory and compare the two but there is no evidence that they
do this. Furthermore, as pointed out earlier, past-tense forms cannot always
be predicted from the phonological form of the stem, since some identical
stems have dierent past-tense forms. Therefore, any network which only rep-
resents phonological form will be unable to learn certain mappings.
Inectional properties such as the choice of a particular ending are asso-
ciated with words, not phonological forms; so to be able to deal with homo-
phones, the processing system must be able to identify unique words. Words
are not just strings of phonemes: they are pairings of a phonological form and
a semantic representation. It would seem, then, that the problem could be
solved simply by giving the system access to semantic information. Such a
solution was actually implemented (on a limited scale) by MacWhinney and
Leinbach (1991; see also Cottrell and Plunkett 1994), who trained their model
on input representations which contained semantic features in addition to
phonological representations. The semantic features allowed the network to
distinguish homophonous stems and to associate them with dierent past-
tense forms. Thus, the network was able to supply rang, ringed and wrung as
the past tense forms of ring1 make a ringing sound, ring2 put a ring around
and wring, respectively.
The critics, however, were not impressed with this solution. Pinker and
Prince (1988: 175), Kim et al. (1994), and others have pointed out that the
choice of past-tense form does not actually depend on semantic properties: all
that matters, as far as the inectional system is concerned, is that ring1, ring2
and wring are dierent words. The rule-inducing mechanism, they argue,
cannot be sensitive to actual meanings, or it would extract incorrect meaning-

based generalisations:
Adding semantic features to a distributed representation has the eect

of dening semantic subtypes of verbs that one would expect to have
similar past-tense forms, just as overlap in phonological features denes
clusters of verbs with similar past-tense forms but this prediction, in
the case of semantic features, is not true. For instance, though slap, hit
and strike have similar meanings, they have dierent past-tense forms
slap has the regular past tense form slapped, hit has the no-change past-
tense form hit, and strike has the stem-changing past-tense form struck.
Conversely, though sting, sing, swing, and so forth form their past tenses
in similar ways, they do not comprise a semantically coherent set of
verbs. This is true not just of these examples, but throughout the verb
lexicon. (Kim et al. 1994: 202)
It is true that the semantic features of the verb are not relevant in English.
But the pattern-extraction mechanism cannot be blind to semantics because
some inectional endings in some languages are associated with classes of
stems sharing certain semantic properties. The choice of the genitive inection
of masculine nouns in Polish, for example, to the extent that it is predictable
at all, depends primarily on semantic factors. Thus, the learning system cannot
simply ignore semantics: it must be able to identify which (if any) semantic fea-
tures are relevant to a particular inection.
But will giving the network access to semantic representations do the trick?
According to Pinker and Prince (1988) and Kim et al. (1994), the answer to
this question is an emphatic no. They point out that the choice of inection
depends not on the presence or absence of a particular semantic property, but
on the grammatical structure of the word. As explained earlier, denominal
verbs, deverbal nouns, bahuvrihi compounds, nouns derived from names and
so on take regular inections in English, even when they are homophonous
with irregulars. Thus, the past tense of spit put on a spit is spitted, not spat,
and the plural forms of sabre-tooth and Mickey Mouse are sabre-tooths and
Mickey Mouses, respectively, not sabre-teeth and Mickey Mice. The reason for
this, according to the dual-mechanism theory, is that these words are exocen-
tric (headless), and so information about irregular inections cannot be
passed on to the derived word. This generalisation, however, cannot be cap-
tured using semantic features, since headlessness is not a semantic property:
no aspect of meaning is shared by spit put on a spit, sabre-tooth, Mickey
Mouse and all other exocentric words. Endocentric compounds, on the other
hand, always take the same inections as their heads, regardless of their
meaning. For example, chessman and strawman pluralise like man, even
though they do not refer to adult male human beings; and become, overcome,
come about, come down with, come on and so on share the same irregularity as
come (their past-tense forms are became, overcame, came down with, came on)
in spite of the fact that they designate very dierent types of events.
Semantic extensions also tend to take the same inection as the basic sense
on which they are based, however tenuous the link between the senses. This is
particularly dramatic in the case of verbs like come, go, have and get, which,
as Pinker and Prince put it, are magnicently polysemous . . . yet they march
in lockstep through the same non-regular paradigms in central and extended
senses regardless of how strained or opaque the metaphor (1988: 11213).
Consider the verb get. The Collins COBUILD English Language Dictionary
lists 43 senses for the verb itself and 94 additional senses for various verbpar-
ticle combinations such as get o, get out and get up. All these senses have the
same past-tense form, got, because they all share the same head.
Kim et al. (1991, 1994) provide experimental evidence suggesting that
English speakers, including young children, are sensitive to the subtle dier-
ence between exocentric and endocentric words. In one experiment, they pre-
sented 35-year-olds with novel verbs homophonous with existing irregular
verbs. Some of the novel verbs were denominal (for example, the experimenter
announced that he was going to y this board, and proceeded to put ies all
over the board), and some were semantic extensions (the experimenter said
that Mickey was going to y down the road and then had him drive very fast
down the road). Subsequently, the children were asked to complete a sentence
describing what had just happened (Mickey just . . .). The general nding was
that children were more likely to use the regular inection with denominal
verbs than with semantic extensions in other words, they were more likely to
say You ied the board than Mickey just ied down the road. However, the
results were very noisy. The children gave the expected regular response to
denominal verbs only 64 per cent of the time, and the expected irregular with
semantic extensions only 23 per cent of the time. They also inappropriately
regularised 47 per cent of the semantic extensions (e.g. Mickey just ied down
the road), and used the irregular form with 6 per cent of the denominal verbs
(e.g. You just ew the board, meaning You just put ies all over the board).
In another experiment, Kim et al. presented nouns with irregular roots as
names (e.g. Mr Tooth), as exocentric compounds (e.g. snaggletooth, a stued
toy with large teeth) or as ordinary nouns (purple tooth), and elicited plural
forms from slightly older children (7- and 8-year-olds). The children used the
expected irregular plural with 94 per cent of the ordinary nouns. Rather unex-
pectedly, however, they also used it with 70 per cent of the names and exocen-
tric compounds. The dierence in frequency of irregular responses is
signicant, showing that the children were sensitive to the distinction between
ordinary nouns and the same roots used as names or in exocentric com-
pounds; but the high proportion of irregular responses in the latter case is
problematic for the dual-mechanism theory.
Kim and colleagues argue that these results show that speakers base their
decision whether to use the regular or the irregular inection on purely gram-
matical properties such as headedness not on semantics. For example, speak-

ers use the regular inection with exocentric compounds like sabre-tooth
because they know they are headless, and the irregular form with extended
senses of irregular words because they know that semantic extensions share
the same head. This solution, however, raises an obvious problem namely,
how speakers determine whether two words share the same head. For example,
how can they tell that tooth is the head in the compound wisdom tooth, but not
in sabre-tooth, or that all the senses of get share the same head? It is clear that
the single most important clue (apart from identical phonological form) is
meaning. A wisdom tooth is a kind of tooth, but a sabre-tooth is a very dier-
ent kind of thing: a tiger. The use of the verb y to mean drive fast is clearly
related to the basic sense (move rapidly through the air), while the novel
denominal sense (to put ies on) is not. With magnicently polysemous
verbs like get, things are slightly more complicated. Pinker and Prince are right
to point out that the various senses of such verbs do not share any semantic
properties (apart from the fact that, like other verbs, they designate relation-
ships that unfold in time). However, there are many local similarities between
the various senses. For example, the obtain sense (as in Hes trying to get an
antique lawnmower) and the receive sense (He thinks he will get an antique
lawnmower for Christmas) share the element of coming to possess; the
become sense (She got pregnant) and the cause sense (He got her pregnant)
both designate events which involve a change of state; and so on. Presumably
learners use such local similarities to make connections between the individ-
ual senses, and eventually integrate them into a single semantic network. The
fact that the various senses share certain grammatical properties (including
the past-tense form) might provide an additional clue.
Thus the decision to use the regular or the irregular inection ultimately
depends on a combination of phonological and semantic factors. A form used
to express a new meaning will inherit the past-tense form from an established
form meaning pairing if the new use is perceived by speakers to be the same
word as the latter. If it is not, the usual conventions of the language apply. In
English, which has only one truly productive pattern for both the past tense
and plural, this almost always means that the regular endings are used.
As pointed out earlier, a word is neither a phonological form nor a seman-
tic representation, but a pairing of the two. For speakers to judge a new
formmeaning pairing to be the same word as an existing irregular, it must
have the same phonological form as the irregular word and share enough rel-
evant semantic properties to be subsumed under the same superordinate cate-
gory as at least one established sense of the word. This explains why words
derived from another category (e.g. a noun derived from a verb or vice versa)
are generally inected regularly in English even if they are homophonous with
existing irregulars. Such a derivation necessarily involves a dramatic change in
meaning, and the derived form and the base cannot be subsumed under the
same semantic category (for example, ring a circular object refers to an
object, while ring to put a ring around designates a kind of action). Because
of this, the new form tends to be perceived as a dierent word and cannot
inherit the grammatical properties of the base; hence, speakers apply whatever
inection would normally apply to the new word.
The above proposal is quite similar to that put forward by proponents of
the dual-mechanism theory. The main dierence is that it explains grammati-
cal behaviour without appealing to an autonomous level of grammatical rep-
resentation. It also has an additional advantage, in that it can explain why
semantic extensions sometimes lose the grammatical properties of the moti-
vating sense and take a dierent inection, and why some irregular-sounding
words derived from another category are inected irregularly.
Situations of the rst kind arise when the extended sense is so dierent from
the motivating sense that the two are no longer perceived as the same word. It
was noted earlier (cf. example (2)) that the genitive form of tenor in Polish is
tenoru when it refers to a singing voice and tenora when it refers to a singer
with that voice. The contrast is quite systematic: words designating other
singing voices also take -u, while words designating singers take -a. In English,
the plural of louse, when used to refer to a group of particularly unpleasant
people, is louses, not lice; and the past tense of hang to execute by hanging is
hanged, not hung. Occasionally, when the link between the basic sense and the
extended sense is not very salient, speakers may be unsure whether or not they
are the same word. When the word mouse refers to a small grey rodent, its
plural form is mice. Many English speakers, however, hesitate to use the irreg-
ular plural when talking about more than one computer mouse, although they
tend to prefer it to mouses. It seems that the reason why they are so squeam-
ish about mice is that the two concepts are perceived as very distant. The com-
puter device and the rodent do share certain properties they are both small
and grey but they are very dierent kinds of things, so many speakers judge
such supercial similarities to be irrelevant. Because the link between the two
concepts is so tenuous, speakers are not sure whether they should be regarded
as the same word, and, consequently, they are not sure which plural form they
should use when talking about several pointing devices.12
Evidence that words derived from another category are not invariably given
regular inections can be found in a series of experiments conducted by
Ramscar (2002). In one experiment, participants were introduced to the nonce
noun frink, which was dened as the Muscovite equivalent of the Spanish
tapas; it is served in bars, and usually comprises chilled vodka and some salted
or pickled sh. They were then presented with a homophonous verb meaning
to consume a frink, and were asked to supply the past-tense form. A substan-
tial majority (73 per cent) produced an irregular form such as frank or frunk.
In another experiment, Ramscar had speakers rate the acceptability of past-
tense forms of real verbs which were either denominal (e.g. sink meaning to
hide in a sink) or semantic extensions of an irregular verb (e.g. the metaphor-
ical sense of sink, as in My hopes sank). Two other groups of participants rated
the same verbs for denominal status and semantic similarity to the prototyp-
ical sense of an existing irregular. Ramscar found that semantic similarity and
grammatical status were highly correlated, and that speakers intuitions about
semantic similarity to the basic sense were a better predictor of the acceptabil-
ity of irregular past-tense forms than their intuitions about denominal
status.13
Thus the data discussed by Pinker, Marcus, Kim and their colleagues does
not necessarily imply that an abstract autonomous level of grammatical struc-
ture is necessary to explain speakers abilities. On the other hand, the facts
need to be accommodated in some way, and it is clear that they cannot be
accommodated in any existing connectionist network. Existing connectionist
models of the past tense do not represent words; and although they are good
at categorising, it is doubtful that they could learn to make the kind of cate-
gorisation judgements required in this task. To deal with the more complex
problems discussed in this section, more sophisticated models will be required
which have access to various types of linguistic and extralinguistic knowledge.
This brings us to the most serious problem yet. All existing connectionist
models are quite restricted in application. The models discussed in this chapter
produce past-tense forms of English verbs. Regiers network, briey described
in Chapter 7, learned to output terms corresponding to various locative rela-
tionships. Other networks had been trained, with varying degrees of success,
to identify word boundaries, assign words to grammatical categories, or
predict the next word in a sentence (Elman 1990, 1993; Gasser and Smith
1998). Thus, each of these models learned to carry out a specic and fairly
simple operation on a particular kind of input. It is not clear, however,
whether such models will scale up that is to say, whether they will be able
to perform more complex linguistic tasks.
More complex types of processing would obviously require more complex
networks. But a more complex network isnt simply a network with more units.
Merely adding units to existing networks is unlikely to solve the problem, for
several reasons. First, as new nodes are added, the number of connections
between them grows exponentially. This means that there are more and more
weights to be set, and as a result, adjusting the weights soon becomes compu-
tationally intractable. Therefore, larger networks will need to show more selec-
tive patterns of interconnectivity, more like the patterns found in the brain,
where there are assemblies of units richly connected with each other, but with
fairly sparse connections with units belonging to other assemblies.
More importantly, the various subtasks involved in language processing
inecting words, identifying word boundaries, accessing phonological forms,
parsing rely on dierent kinds of knowledge, and hence, in network terms,
require dierent connection strengths. A network trained to produce past-
tense forms cannot identify word boundaries, and vice versa. Training it to
perform the second task would involve altering connection weights, which
would interfere with performance on the rst task. Thus, if a network is to be
able to perform a number of dierent tasks, various subparts will have to spe-
cialise. Complex problems may also require decomposition into smaller tasks,
and these, too, would have to be assigned to dierent subnetworks.
It follows that a more realistic model of human linguistic abilities will
require a certain degree of functional specialisation or modularisation.14
Unfortunately, it is not clear how such functional specialisation could be
achieved in networks. Presumably some architectural features will have to be
built in (as in Regiers model), but many leading gures in the eld believe that
functional specialisation could arise as a result of learning in other words,
that future models will be able to self-organise (see Elman et al. 1996).
All these problems notwithstanding, it is undeniable that connectionist
models have stimulated research and helped us to develop new ways of think-
ing about language development and language processing. It is important to
realise, however, that connectionism is not a theory of language processing or
development: it is an approach to modelling these and other phenomena.
Connectionist networks can be used to model learning in a variety of frame-
works, empiricist, nativist or constructivist. Specic models, of course, do
embody assumptions about how language is represented in the brain; however,
the theories embodied in existing networks are very simple, almost trivially
simple. The networks developed so far are useful research tools, but they are
grossly inadequate as models of human linguistic abilities.
This does not mean that more realistic models are impossible in principle.
Connectionism is a relatively young endeavour, and much of the emphasis so
far has been on the mechanics of modelling developing suitable algorithms,
architectures, training regimes, and methods of analysing network behaviour.
Future models, it is to be hoped, will implement more sophisticated theories
of language and language acquisition.
Notes
1. Although applicability and frequency often co-vary, they are logically

independent. One can envisage a system in which some inections have
narrow applicability (e.g. apply only to stems with specic phonological
properties) but have high type frequency because these phonological
properties are typical of most stems in the language.
2. The frequency of -s overgeneralisations appears to increase with age,
however. They accounted for 58.5 per cent of all errors in an elicited pro-
duction study with 39-year-olds conducted by Clahsen (1999).
3. Clahsen et al. assume a model of morphology proposed by Kiparsky
(1985), according to which the morphological component is divided into
several ordered levels, and processes that apply at later levels cannot alter
units assembled earlier. According to the model, irregular inections
apply at level 1, compounding at level 2, and regular inections at level 3.
Because irregular inections are added at level 1, before compounding,

they can occur inside compounds; but regular inections, which apply
after compounding, can only be added to the compound as a whole, not
to an element inside a compound.
4. For an explanation in an optimality theory framework, see Wegener
(1999).
5. The distribution of the two variants is determined by very general phono-
tactic rules, and some linguists (e.g. Nagrko 1998) even consider them
allophones of the same phoneme. This decision, however, depends on
assumptions about the phonemic status of preceding consonants which
are somewhat controversial.
6. Feminine nouns usually end in -a, masculines in a consonant, and neuters
in -o, -e or -e.
7. These morphological eects are not independent of semantics. Nouns
derived using the a-friendly suxes -acz, -arz, -ak, -nik, -ec typically des-
ignate animate beings. The -a ending is also preferred with nouns desig-
nating small objects, and hence diminutives take -a. On the other hand,
-izm and -ot, two of the suxes which are associated with -u, are used to
derive nouns designating ideologies and noises, i.e. abstract concepts. The
phonological associations between axes and stem-nal consonants may
result from a similar trickle-down eect (see Janda 1996; Kottum 1981).
8. Some masculine nouns (e.g. syk hiss, pisk squeak, plusk splash, klekot
rattle, stukot clatter) might be regarded as onomatopoeic in that their
phonological form resembles the noises they designate. These, like other
masculine nouns designating sounds and abstract concepts in general,
require -u. However, they are ordinary noun roots rather than actual ren-
ditions of sound, which is what Marcus et al. seem to have in mind, so
they are not considered here.
9. The gender of the nonce words used in the experiment could be predicted
from the phonological form of the nominative as well as from morpho-
logical cues such as agreeing adjectives and/or demonstratives.
10. The acquisition of the genitive by these children is described in more
detail in Dabrowska (2001, 2004). For further details about the corpus
from which the data was drawn, see Smoczynska (1998).
11. Note that this problem did not arise in the genitive study, for similar
reasons: in the genitive, the masculine class is split up between -a and -u.
12. Actually, usage is somewhat variable in all four cases. For example, some
English speakers use the irregular inection with both senses of hang
(suspend in the air and execute by hanging), presumably because both
refer to an act of suspending something so that it does not touch the
ground.
13. This nding contradicts earlier research by Kim et al. (1991); however,
there are some serious methodological problems with this study. See
Ramscar (2002) for a discussion.
14. Note that the term modularisation is used here to refer to functional spe-
cialisation. The resulting modules, or subnetworks, are not necessarily
anatomically distinct. Moreover, they receive input from other subnet-
works, so they are not informationally encapsulated.
9 Syntactic constructions
We saw in the previous chapter that there is no evidence for a strong dissocia-
tion between grammar and the lexicon in morphology: in principle, the same
mental mechanism can account for speakers ability to supply both regular
and irregular inections. It was also suggested that speakers extract patterns
at various degrees of abstraction and that associative memory plays a promi-
nent role in this process.
In this chapter, I will argue that the same may also be true of syntax. This
is seen most clearly in acquisition, so the main body of the chapter will be
devoted to discussing developmental data. We will begin with a summary of
studies showing that there is a very close relationship between lexical and
grammatical development, at least in the early stages of language acquisi-
tion. We will then look at research which might explain this relationship, and
examine converging evidence from a variety of sources suggesting that the
starting point of syntactic development is the acquisition of a repertoire of
rote-learned formulas, or multi-word phrases paired with meaning. Finally,
we will look in considerable detail at the development of a particular con-
struction (or rather family of constructions) English interrogatives in
order to see how more creative patterns develop from such formulas, and
conclude with some observations about lexically specic units in adult lan-
guage.
1. Ties between lexical and grammatical knowledge
A series of large-scale studies conducted by Elizabeth Bates and her team have
revealed the existence of close ties between expressive vocabulary size and
grammatical development. Bates et al. (1988), in a longitudinal study of 27
children who were followed from the age of 10 months to 28 months, showed
that there was a very high positive correlation (+.83) between expressive
vocabulary size at 20 months and MLU (mean length of utterance) at 28
months. Indeed, the correlation coecient is as high as the correlations
between MLU measures obtained from dierent samples taken from the same
child at 28 months: in other words, vocabulary at 20 months and MLU at 28

months are statistically indistinguishable.
These ndings were conrmed by a second, much larger cross-sectional
study of 1,800 children aged from 16 months to 30 months (Bates, Dale and
Thal 1995; Fenson et al. 1994) which used the MacArthur Communicative
Development Inventory (an extensive questionnaire completed by the parents).
Although this study used a dierent measure of grammatical development (a
grammatical complexity score derived from parental report about the struc-
tures used by their children), the relationship between vocabulary size and sen-
tence complexity was virtually identical (+.84). The same relationship was also
found in various atypical populations, including early and late talkers, children
with focal brain injury, and individuals with Williamss syndrome (WS) (Bates
and Goodman 1999), as well as in normally developing Italian children (Caselli
et al. 1999).
Such correlations cannot be explained simply in terms of maturation.
Although both vocabulary size and grammatical complexity increase with age,
age is a poor predictor of grammatical development. In the large cross-sec-
tional study referred to above, age and vocabulary size together accounted for
71.4 per cent of the variance in grammatical complexity; when age was con-
trolled for, vocabulary size accounted for 32.3 per cent of the variance; but
when vocabulary size was controlled for, age accounted for only 0.8 per cent
of the variance (Bates and Goodman 1999: 46).
Bates and Goodman (1997, 1999) also note there is evidence of close ties
between lexical and grammatical knowledge in language impairment, includ-
ing impairments acquired in maturity, which suggests that these ties are not
restricted to the early stages of language acquisition. All aphasic patients, they
point out, whatever other problems they might have, also have lexical decits:
in other words, there are no impairments which involve grammar and
grammar only. Moreover, there are parallels between patients lexical and
grammatical diculties. In some decits (Brocas aphasia, Down syndrome,
some forms of SLI), the underlying problem seems to be in accessing linguis-
tic representations, and it manifests itself in word-nding diculties and
omission of obligatory function words and inections. Other groups
(Wernickes aphasics, people with WS) make errors of commission rather than
omission: they use both grammatical morphemes and words inappropriately.
Still others (e.g. anomic aphasics and patients with Alzheimers dementia) do
not make outright errors, but have very restricted linguistic repertoires: that is
to say, they are overly dependent on pronouns and other semantically light
forms and the most basic grammatical forms and constructions (see Chapter
4 for further discussion).
How can we explain these close ties between lexical and grammatical knowl-
edge? The key to understanding this relationship, I will argue in this chapter,
is formulas or multi-word prefabs: stored phrases associated with a specic
meaning. Such prefabs are like words, in that they are formmeaning pairings.
Syntactic constructions 161
On the other hand, they are also grammatical constructions: they have inter-
nal structure, and they may be partly underspecied and hence allow some
variation. We saw in Chapter 2 that the use of prefabs is one of the processing
shortcuts that help adults produce and understand complex utterances in real
time. I will now review evidence that children learn such chunks, and that they
play an important role in acquisition.
2. Multi-word units in acquisition
2.1 Premature usage

Most children occasionally say things which appear to be well in advance of
their productive abilities. Nelson reports that the children she studied used
invariant formulas such as want a drink of water, how are you?, youre kidding,
containing the correct grammatical morphemes, long before they used these
morphemes in any other contexts (1973: 1078). Braines (1976) subject David
repeatedly used the phrase Can I (have) X? to make requests when he was still
in the two-word stage, before he used other auxiliaries or learned how to form
yes/no questions. Many children ask formulaic WH questions such as Whats
this?, Wheres X?, Whatchadoing?, often in highly stereotypical contexts,
months before they produce any other interrogative constructions (see below;
also Brown 1968; Pinker 1984; Radford 1990).
Richards (1990) provides an even more striking example. Two children in
his study produced well-formed tag questions at a very early stage, before they
had learned to form yes/no questions or to use auxiliaries as pro-forms. In
some cases, the tag was the rst attested usage of a particular auxiliary. Since
tags are one of the most complex syntactic contexts for auxiliaries, their early
appearance suggests that these two children had learned them as unanalysed
units. This is conrmed by errors such as They swimming intit? (109), I know
him arent I? (77), where the auxiliaries (and in the rst example, also the
subject) in the tag dont match the corresponding units in the main clause.
2.2 Developmental U-curves

In some cases, early correct but restricted usage is followed by periods of
apparent regression, during which the form in question is either absent from
the childs speech or is used inappropriately, before it reappears again, more
reliably and in a wider range of contexts. For example, Bates et al. (1988),
Hickey (1993), Peters (1983) and others have observed that some children
produce functors in their early speech, then leave them out, and then begin to
supply them again more reliably and in a wider range of contexts. In the devel-
opment of WH questions, early correct forms with a contracted copula
(Whats this?, Wheres Daddy?) may be followed by uninverted forms (What
this is?, Where Daddy is?) and then questions with the full form of the copula
in the correct position (What is this?, Where is Daddy?) and eventually ques-
tions with both the contracted and the uncontracted form. Richards (1979)
and MacWhinney (1985) note that while 3-year-old children produce correctly
ordered strings of adjectives before nouns, performance often dips around the
age of 5 before improving again in the school years. The early correct perfor-
mance is attributable to the use of memorised strings such as great big and nice
little.
2.3 Inappropriate and ungrammatical usage

Ungrammatical utterances produced by children are often regarded as evi-
dence that they are relying on their own rules rather than merely imitating
adult models. However, ungrammatical utterances are just as likely to arise
when the child attempts to combine rote-learned chunks, as in the following
examples:
(1) a. I dont wanna dont throw. (Iwamura 1979, cited in Peters 1983)
b. Can I have break it? [wants to break up a tinker-toy contraction]
(Braine 1976)
c. I dont know wheres Emma gone. (Clark 1974)
d. I want I eat apple. (Clark 1974)
e. Wheres the boy brought the milk? [looking for the milk the boy
had brought] (Clark 1974)
In (1a), the child simply incorporated the previous speakers utterance (dont
throw) into the frame I dont wanna X. The next two utterances are constructed
by combining two phrase-sized units from the childs own repertoire. In (1b),
the general-purpose request formula Can I have X? is combined with what is
probably an unanalysed chunk (at the stage when the utterance was produced,
the child used the verb break only in the expression break it). In (1c), the noun
Emma is used to ll the slot in a well-established formula (Wheres X gone?),
and the resulting expression is combined with another formula I dont know X.
Utterances (1d) and (1e) were probably assembled by embedding a simple con-
structed utterance (I eat apple, the boy brought the milk) inside a formulaic
frame (I want X, Wheres X?).
In other cases, an utterance may be grammatical and its formulaic status is
only revealed when it is used in inappropriate circumstances. One of the rst
plural forms used by Clarks son Adam occurred in the sentence Take the cups
away, which he heard in the day nursery. He subsequently used it at home even
when there was only one cup. Her other son Ivan produced the utterance It
will hurt me after a door had slammed in front of him (Clark 1977). Clearly,
he was not using will to signal the future, but simply recycling an expression
he had heard earlier in a similar context. What is noteworthy about this par-
ticular example is that the recycling did involve a certain amount of modica-
tion of the original utterance: he changed the pronoun to reect the change of
speaker.
2.4 Pronoun reversals

This brings us to a particularly interesting usage error in early child language:
pronoun reversals, where the child uses the second person pronoun to refer to
himself, or, less frequently, the rst person pronoun to refer to the interlocu-
tor, as in the following examples:
(2) a. I carry you. [wants to be carried by his father] (Clark 1974)

b. Sit my knee. [wants to sit on the adults knee] (Clark 1974)
c. Lets see you climb up on this one. [said while climbing a low wall]
(Snow 1983)
d. The pjs you got on are slippery. [referring to his own pyjamas]
(Snow 1983)
Such errors have been reported quite frequently in the literature (Chiat 1986;
Clark 1974; Dale and Crain-Thoreson 1993; MacWhinney 1985; Petitto 1987;
Snow 1981), and there is no shortage of theories attempting to explain them.
One popular explanation is that reversals are a result of young childrens
inability to assume the point of view of another person; others have attributed
such errors to the unusual nature of deictic pronouns (unlike most other lin-
guistic expressions, they refer to roles, not categories or individuals), charac-
teristics of the input (in speech addressed to the child, you always refers to the
child and me to the caretaker, so unless the child is exposed to language
addressed to third parties, she has no evidence that pronouns refer to roles
rather than individuals), and to processing complexity (pronoun reversal is a
performance phenomenon, particularly likely when processing demands are
high).
All four factors probably play a role, and their relative importance may vary
over time and from child to child. However, I would like to suggest that the
single most important cause of pronoun reversals is the childs use of precon-
structed phrases.
Several lines of evidence point in this direction. First, cases of consistent
reversals are rare. Most children who reverse pronouns produce correct forms
most of the time, and comprehension tends to be good even in children who
make errors in production (Chiat 1982). This argues against the rst three
hypotheses. Secondly, reversals are particularly frequent in imitations. All the
utterances in (2) above are (deferred) imitations of parental utterances. (2a)
and (2b) are reduced renditions of the parents Ill carry you and sit on my knee;
(2c) is an exact repetition of an utterance used on a number of similar occa-
sions by the childs father; and (2d), an exact imitation of an utterance the
child heard in a very similar context the previous day (he was standing at the
top of the stairs in the morning in pyjamas with fabric feet; he reproduced
the utterance the following morning while walking down the stairs, wearing
the same pyjamas). In a study of deictic pronoun use in linguistically preco-
cious children, Dale and Crain-Thoreson (1993) found that 52 per cent of all
reversals occurred in imitations, in spite of the fact that imitations made up
only 5 per cent of the childrens utterances. Many of the remaining utterances
with reversals could well be delayed imitations which werent recognised as
such because the original parental utterance was not captured on tape. For
example, it is likely that the children whose utterances are reproduced in (3)
below heard a parent say something like Do you want mummy to help you
now?/Should mummy help you now? and Do you want me to help you with this?
(3) a. Mama help you now.

Mama help me now.
b. Want help you with screwdriver.
I want you to help me with the screwdriver.
Dale and Crain-Thoreson argue that reversals result from processing overload
(the child fails to perform deictic shift in cognitively demanding situations),
and note that they are particularly likely to occur in syntactically complex sen-
tences. But if cognitive overload were the problem, reversals should be less fre-
quent in imitations, which require less processing capacity than productive
usage. A more plausible explanation, therefore, is that reversals occur when
children recycle an adult utterance, or parts of it. This often enables them to
produce more complex utterances, but may lead to reversal errors when the
adult utterance contains pronouns.1
It is worth noting that pronoun reversals are especially common in the
speech of two clinical groups blind and autistic children. Autistic children
are well-known for their use of unanalysed formulas, and so this aspect of
their speech supports the claim that pronoun errors are a tell-tale sign of holis-
tic processing. In the case of blind children, inappropriate use of pronouns
may be a consequence of a delayed development of a sense of self (see
Fraiberg 1977) or of limited access to contextual information.2 On the other
hand, there is also evidence that blind children rely on formulaic phrases to a
greater extent than seeing children (Dunlea 1989; Mulford 1988; Prez-Pereira
1994; Urwin 1984; see also Chapter 3), so this could also contribute to the
higher incidence of pronoun errors.
2.5 Filler syllables

Another characteristic feature of formulaic usage is the occurrence of phono-
logical llers unstressed syllables (usually a schwa, a nasalised schwa, or a
nasal consonant) which usually occur in those positions where function words
would appear in adult speech (in English, before open-class words). The fol-
lowing utterances produced by Seth, a child studied by Ann Peters (Peters
1995; Peters and Menn 1993), are a good example (the upper-case F in the
glosses marks the position of the ller):
(4) a. /n si z ht/ F see F hedge.

b. /n si sons/ F see stones?
c. /n si bak/ F see F bark (of tree)?
d. /m pk fawis/ F pick F owers? (Peters 1995: 472)
Seth began using ller syllables while still in the one-word stage. During the
period from 19 months to 21 months, they comprised about 30 per cent of the
morphs in his speech (Peters and Menn 1993: 748). It is often dicult to iden-
tify a specic target for these llers, but their presence lent the entire utterance
a more adult-like prosodic contour. Gradually, the llers gained more pho-
netic substance and became more and more dierentiated and eventually they
evolved into various adult forms.
Peterss interpretation of the data was that Seth was unable to reproduce the
adult utterance in full phonetic detail and attempted to approximate it at the
prosodic level:
In English closed-class morphemes tend to be (parts of) unstressed syl-

lables whereas open-class words generally contain at least one stressed
syllable . . . These two kinds of constituents therefore participate in com-
plementary ways in the rhythm of language, with the open-class items
providing the strong beats and the closed-class items contributing to the
weak ones. Formulaic children seem to be sensitive to both rhythm and
intonation, making noticeable eorts to reproduce them even in their
early utterances. (Peters 1995: 476)
Another possibility is that llers are a kind of all-purpose function word

which is the common denominator of all the function words that occur in a
particular position.3 For example, the prenominal ller occurs in the slot
which is occupied by determiners in adult language and its phonetic indistinct-
ness could be a result of the fact that it captures properties shared by several
dierent words: a, an, the etc. Seth knows that he should use one of these
words, but he is not sure which one so he uses a time-honoured strategy by
no means conned to children learning language: when in doubt, mumble.
Whatever the reason for the underspecication of ller syllables, they
suggest that Seth is working with units larger than a single word formulas
such as NF-see-F-THING or possibly even NF-PROCESS-F-THING (where
NF = nasal ller and F = ller), each associated with a specic intonational
contour. The llers are not nonsense syllables inserted between content words:
they actually help to dene the frame in which other items occur. They also
act as placeholders for grammatical morphemes which will emerge later, and
the frame itself embodies the childs emergent knowledge about the distribu-
tion of these morphemes.
How typical (or untypical) was Seth? Is there any evidence that other chil-
dren use this strategy as well? The use of ller syllables has been reported by
quite a few researchers working on the acquisition of a variety of languages
(Peters 2001); but it is likely that most cases go unnoticed, partly because llers
are most characteristic of holistic children (and holistic children tend to be
underrepresented in language-acquisition research), and partly because
researchers often ignore syllables which have no discernible function. For
example, Braine (1976: 43) notes that one of his subjects inserted short
unstressed syllables (usually [] or [d]) before words, both utterance initially
and medially. These syllables didnt seem to make any semantic contribution
to the utterance and occurred before verbs as well as before nouns, so they
could not be determiners. Because of this, Braine decided to ignore them.
Presumably others have done likewise. Since many researchers working on
grammatical development use orthographic transcriptions, they may not even
be aware of the presence of llers in the speech of the children they study
unless they have transcribed it themselves.4
2.6 Lexically based patterns

Young childrens grammatical knowledge can be surprisingly specic in
many cases, it is restricted to the combinatorial possibilities of a particular
word or a narrow class of words. Some of the best-known evidence for this
comes from studies of rst word combinations by Braine (1963, 1976), who
observed that his 2-year-old subjects reliably put some words in the correct
positions in the utterance, but positioned other words belonging to the same
grammatical category inconsistently. For example, one child consistently put
colour words and the adjectives big, little, hot before the noun, but positioned
(all) wet sometimes before and sometimes after the noun. This suggests that
he was working with fairly specic positional patterns, or formulaic frames,
such as little + X, hot + X, COLOUR WORD + X, rather than a fully general
rule stating that adjectives come before nouns. Likewise, Braine found no evi-
dence that 2-year-olds had a general rule for ordering the verb and the direct
object: both VO and OV orders occurred, though with dierent verbs. Thus,
the children seemed to be using very specic patterns like want + X, see + X,
eat + X, drink + X and so on.
Similar ndings were reported by Bowerman (1976), Tomasello (1992) and
others. Between the ages of 16 months and 18 months, Tomasellos daughter
Travis used 23 verbs and verb-like words (out of a total repertoire of 45) in
combination with other words. Of these, 21 were consistently ordered. For
example, Travis always put the object looked for after the amalgam nd-it,
producing utterances such as nd-it bird, nd-it chess, nd-it brick, nd-it ball
etc. On the other hand, the object of get-it occurred both before and after the
verb: block get-it, bottle get-it, get-it hat, get-it birds.
The existence of such lexically specic patterns in the speech of young chil-
dren is well documented. What is less well known is that they persist for a long
time after the child has advanced beyond the two-word stage, and seem to play
an important role in the transition to multi-word speech. Thus, when Travis
began to produce utterances containing three or more words, they were nearly
always constructed by adding a single element to an existing formula or
embedding one formula inside another. For example, adding an actor nominal
to her ride X formula allowed her to produce utterances such as Big Bird ride
horsie; embedding Daddys X in see X resulted in novel expressions such as see
Daddys car; big X plus X stuck yielded big rock stuck; and so on.
Several facts about Traviss usage at this time suggest that her grammatical
knowledge remained verb-specic at least until age 2;0. She tended to use
dierent combinations of arguments with dierent verbs. Many verbs were
combined with one argument only: some only with an agent, others only with
a patient or only with a location. This cannot be explained away as a process-
ing limitation, since other verbs occurred in frames with two or more argu-
ments: for example, she consistently expressed all three arguments with the
verb gave (but, interestingly, not with give, suggesting that the past- and
present-tense forms were independent constructions). What is more, she expli-
citly marked locative relations in utterances with some verbs but not others:
for example, she would say Travis sit down chair (without a locative preposi-
tion) but Dana push me real high in a bagswing. Her use of verbal inections
was also restricted to specic verbs. She used 24 verbs in the past-tense form
and 23 in the progressive, but only four verbs were used with both endings. Her
total verb vocabulary at the time was over 150, so over two-thirds of her verbs
were never inected (Tomasello 1992: 254). Likewise, explicit agreement
marking was restricted to the copula and a few other verbs.
Tomasellos diary method allowed him to chart the emergence and gradual
development of his daughters formulas in ne detail, but it is dicult to draw
general conclusions from a single case study. How typical was Travis? Are
other children equally restrained in their generalisations, or do they form
abstract rules right from the beginning? Fortunately, we do have similar data
from other children which allow us to answer these questions.
Lieven et al. (1997) studied the distributional regularities in the speech of
11 children. They recorded all multi-word utterances until the child reached
25 word-based patterns. At this point, the childrens ages ranged from 1;7 to
2;9 and their MLUs from 1.41 to 3.75 (mean 2.41). The researchers then
divided all multi-word utterances into three categories: constructed (lexical
formulas which had occurred at least twice before), intermediate (the second
occurrence of a particular positional pattern, or the rst if both words had
been attested in single-word utterances), and frozen phrases (utterances which
contained at least one word which did not occur in any other expression). They
found that lexical formulas accounted for 60.3 per cent of all utterances (range
51.371.9 per cent). A further 31.2 per cent were frozen phrases. Of the
remaining 8.4 per cent, about half showed no positional consistency, and the
other half did (and hence could be instantiations of emerging lexical frames).
The more complex sentences that children produce also tend to be quite for-
mulaic. A detailed analysis of utterances with nite complement clauses pro-
duced by children aged from 2 to 5 conducted by Diessel and Tomasello (2001)
revealed that the main clauses in such sentences are usually very short.
Typically, they consist only of a pronoun (I or sometimes you) and a present-
tense verb, with no auxiliaries or adverbials; and even the subject is sometimes
omitted. Diessel and Tomasello point out that the proposition in such utter-
ances is typically expressed by the complement clause, while the main clause
functions as an epistemic marker (I think, I bet), a deontic modality marker (I
hope, I wish) or an attention getter (see, look), so the structures are, in eect,
monoclausal, and can be paraphrased with a simple sentence:
(5) a. I think its in here. [= Maybe its in here.]

b. I hope he wont bother you. [= Hopefully he wont bother you.]
This analysis is supported by the fact that what appears to be the subordinate
clause is virtually never introduced by a complementizer, and that the main
clause sometimes comes after the complement clause (as in Its in here, I
think).
The children in the study also produced some sentences which do appear to
involve true subordination. These sentences were much less formulaic: the
main clauses contained a variety of subjects and verb forms as well as modi-
ers and expressed an independent proposition, and the subordinate clauses
were sometimes introduced by complementizers. It is noteworthy, however,
that the vast majority of them had one of just three verbs say, tell, pretend
as the main predicate. This, argue Diessel and Tomasello, suggests that they
were licensed by fairly specic schemas like Im just pretending S, NP SAYS,
NP TELL NP S (where SAY and TELL stand for various forms of these two
verbs) rather than a fully general subordination rule, since if children had a
general rule, they would apply it to sentences containing other verbs as well.
Lexically specic units, then, are a ubiquitous feature of early production,
which strongly suggests that young childrens knowledge may be described in
lexically specic terms. However, it is also possible that children do, in fact,
have fully general, adult-like rules but for some reason do not use them in pro-
duction. In order to rule out this alternative possibility, Tomasello and col-
leagues have conducted a series of experiments in which they investigated
childrens ability to use unfamiliar words in novel ways. They found that even
children as young as 1;6 to 1;11 freely assimilate novel nouns to a variety of
constructions. Thus, if they are taught the nonce word wug referring to an
unfamiliar animal, they spontaneously use it in combination with items which
are already part of their repertoire, producing utterances such as wheres the
wug?, I want my wug, wug did it and so on (Tomasello et al. 1997). However,
childrens performance in experiments involving novel verbs is dramatically
dierent: the under-3s tend to use novel verbs only in the construction in
which they heard them. For example, Brooks and Tomasello (1999) taught
children aged about 2;11 two novel verbs designating highly transitive actions.
One of the verbs was presented in the transitive construction (Big Bird tammed
the car) and the other in a passive construction (The car got meeked by Big
Bird). They then attempted to elicit transitive sentences by asking what the
agent was doing. They found that only 25 per cent of the children who heard
passive models were able to use the novel verb in an active transitive sentence
describing the same action with a dierent set of participants. However, when
tested on the verb they heard in the transitive construction, all of the children
succeeded in this task.
In other experiments, Tomasellos team attempted to elicit a full transitive
with a novel verb modelled in intransitive (The lion is gorping), imperative
(Gorp, lion!) and presentative constructions (Look, this is called gorping). The
results are very consistent: children aged 2;0 and below virtually never produce
full transitives with a novel verb modelled in another construction; the pro-
portion of children who succeed in such tasks increases steadily with age until
it reaches 80 per cent by about 4;0; after 4, progress levels o slightly, but by
about 8 years of age, virtually all children are able to use a novel verb in the
transitive (see Tomasello 2000). Berman (1993) obtained comparable results
with Hebrew-speaking children. In her experiment, the novel verb was mod-
elled in the intransitive construction and elicited in the transitive. In Hebrew,
this transformation is more complex, as it requires morphological changes to
the verb as well as a rearrangement of the arguments, and so, not surprisingly,
the task turned out to be even more dicult for children: only 9 per cent of
the youngest participants (2;9) were successful; the proportion rose to 38 per
cent by 3;9 and to 69 per cent by 8;0.
While most of the work with novel verbs focused on the transitive construc-
tion, the same methodology has also been used to study childrens productiv-
ity with other structures, with similar results. Brooks and Tomasello (1999),
for example, also attempted to elicit passives, and found that 85 per cent of the
children in the youngest group they tested (age 2;11) were able to produce a
full passive with a novel verb if the verb had been modelled in the passive (with
dierent NPs expressing the agent and patient roles). However, if they had
only heard the verb in active sentences, none of the children produced a full
passive, and only 12 per cent were able to produce a truncated passive (a
passive clause without the by-phrase).
The Brooks and Tomasello study also provides evidence bearing on another
acquisition issue. English-speaking children begin to produce full passives
fairly late typically between 4 and 5 years of age. One explanation for their
late appearance, often put forward by researchers working in the generative
tradition (e.g. Borer and Wexler 1987), is that passives are dicult because
they involve movement, which is a late-maturing syntactic ability. An alterna-
tive account, preferred in usage-based approaches, attributes childrens di-
culties to the fact that full passives are very rare in the input: Gordon and
Chafetz (1990), for example, found only four tokens of this construction in
over 86,000 child-directed utterances. The fact that most of the under-3s in the
Brooks and Tomasello study were able to use full passives productively after
brief but intensive training (two 30-minute sessions containing about 100
models) suggests that the latter explanation is correct: that is to say, that fre-
quency in the input is the determining factor.
Further support for the idea that young childrens knowledge about word
order is lexically specic rather than general comes from comprehension
studies involving nonce words. In one such study (Akhtar and Tomasello
1997), children were taught a new activity involving two characters and an
apparatus; then they were given two new characters and the apparatus and
asked to make Cookie Monster dack Big Bird. The main nding was that while
older children (mean age 3;8) did very well, only three out of ten 2;9-year-olds
performed above chance.
But perhaps the most striking evidence for lexical specicity comes from
two studies in which novel words were presented in constructions with non-
canonical (i.e. non-English) word order. In the rst of these, conducted by
Nameera Akhtar (1999), children were taught novel verbs designating highly
transitive actions in syntactic frames with SOV or VSO word order (e.g. Elmo
the car dacking, Gopping Elmo the car) and then encouraged to describe scenes
involving dierent participants using the newly learned verbs. The question
was whether they would correct to the canonical SVO pattern (thus demon-
strating verb-general knowledge of word order) or use the unusual word order
they heard in the model sentences. Akhtar found that the oldest children
(mean age 4;4) virtually always used the SVO pattern, no matter what word
order was used in the model sentences, while the younger children (ages 2;8
and 3;6) corrected to the canonical word order or used the unusual construc-
tions they had been to exposed to during the experiment equally frequently.
However, all age groups consistently corrected to the canonical word order
when a familiar verb was presented in an SOV or VSO pattern (for example,
they would never say He the car pushing). Thus, the younger childrens use of
the unusual word order cannot be explained by appealing to irrelevant factors
such as reluctance to correct an adult. Another interesting nding was that in
nearly all the utterances with non-canonical orders, the agent and patient
arguments were expressed by lexical NPs rather than pronouns, whereas half
of the utterances that were corrected contained pronouns: that is to say, the
younger children would happily say Ernie the spoon dacking but not He it
dacking. This suggests that the corrected utterances may also have been pro-
duced using lexically specic schemas based on pronouns (such as Hes
VERBing it).
The second study, by Abbot-Smith et al. (2001), involved slightly younger

children (aged 2;4 and 3;9) and used intransitive action verbs (i.e. unergatives)
rather than transitives. The verbs were presented in three conditions:
1. Novel SV: novel verb in the canonical order, e.g. The cow bang;
2. Novel VS: novel verb in a non-canonical order, e.g. Bang the cow
(meaning The cow is wobbling out);
3. Familiar VS: familiar verb in non-canonical order, e.g. Jumping the cow
(The cow is jumping).
The results were broadly compatible with those of the previous study: the 3-
year-olds consistently used SV order in all three conditions, while the 2-year-
olds generally corrected to the canonical word order with familiar verbs (72
per cent of the time), but not with novel verbs (only 21 per cent of the time),
showing that they lacked a general SV schema they could rely on in produc-
tion. As in the Akhtar study, the non-canonical word order was virtually never
used with pronouns (i.e. children did not produce utterances such as Baed
she), while the majority of the utterances with corrected word order contained
pronouns; the 2-year-olds, in fact, were almost ve times more likely to correct
to the canonical word order when the subject was a pronoun. Also, when using
the non-canonical VS order the children never added the missing auxiliary, but
they did tend to add it when using the SV order. Interestingly, they were more
likely to add it after a pronominal subject than after a lexical subject (70 per
cent v. 26 per cent of the time in the 2-year-old group). These asymmetries
suggest that they have learned multi-word units such as Hes VERBing and
Shes VERBing.
There is, however, an interesting twist to these weird word order studies. In
both experiments, children were considerably more likely to reproduce the word
order they heard with a particular nonce word if it was compatible with the
canonical word order in English. Akhtars children produced approximately
four times as many sentences with novel verbs in the control SVO condition as
in the experimental conditions. Abbot-Smith et al. also recorded about four
times as many productive uses of the nonce verb in the Novel SV condition as
in the Novel VS condition. One explanation for this pattern of results is that the
children were actively avoiding the strange word order used by the experimenter,
which would suggest that they had some nascent verb-general knowledge, even
if it was too weak to be reliably used in production. Another possibility is that
knowing a large number of verbs associated with a particular construction
makes it easier to acquire new verbs occurring in the same construction. This
would entail that children make connections between their lexically specic pat-
terns from the very beginning, although they may need time and/or a critical
mass of exemplars before they extract a general rule.
Further evidence about the nature of young childrens syntactic representa-
tions comes from studies which used the preferential looking paradigm. This
method involves presenting the child simultaneously with two videos depict-
ing two dierent scenes (e.g. a causative action such as a duck lifting a rabbits
leg and a non-causative action such as a duck and a rabbit each covering its
own eyes) and a test sentence which describes one of them (e.g. The duck is
gorping the bunny). The experimenter then measures how long the children
look at each screen. If they spend more time looking at the matching video, it
is concluded that they are able to extract some information from the syntactic
form of the construction. The greatest advantage of this method is that it can
be used with very young children (1314 months); the main disadvantage is
that it is not clear what kind of information children extract from the linguis-
tic stimuli (cf. Naigles 1990; Tomasello and Abbot-Smith 2002).
Using the preferential looking procedure, Hirsh-Pasek and Golinko (1996:
experiments 5 and 6) found that the oldest children in their study (mean age
289 months) looked longer at the video showing a causative event when pre-
sented with a novel verb in a transitive sentence, but performed at chance when
presented with intransitive sentence; the middle group (24 months) showed a
preference for the matching video with familiar verbs only, thus demonstrating
verb-specic knowledge; and the youngest group (19 months) showed no evi-
dence of understanding the sentential context with either familiar or novel
verbs. Kidd et al. (2001), who tested slightly older children (mean age 30
months), found that they performed above chance on both transitive and
intransitive sentences with nonce words. Another study (Naigles 1990) suggests
that the relevant knowledge may appear somewhat earlier, around 25 months.
Thus, the preferential looking ndings are broadly compatible with the pro-
duction studies described earlier, in that they suggest that lexically specic
knowledge precedes verb-general knowledge; they do, however, suggest that
verb-general knowledge emerges considerably earlier. It is important, however,
not to overinterpret these ndings. The fact that the children were inuenced
by the linguistic form of the sentences accompanying the video presentation
shows that they were picking up on some property or properties of the sen-
tences and linking them to the display but it is by no means clear what exactly
they were attending to. Specically, there is no strong evidence that they were
using word order to assign agent and patient roles: it is possible that they asso-
ciated the experimental sentences with fairly holistic representations of the cor-
responding visual displays. For instance, they could have linked transitive
utterances such as The duck is gorping the bunny to scenes involving two par-
ticipants doing dierent things, and intransitive sentences such as The duck and
the bunny are gorping to scenes in which the two participants do the same thing;
note that in the latter case, the relevant cue might not be word order but the
word and.
Further evidence suggesting that young childrens constructions are lexi-
cally specic while older children have more abstract representations comes
from an experiment by Savage et al. (submitted). In this study, children were
shown pairs of pictures. The experimenter described the rst picture using a
particular construction (for example, a passive such as The car got smashed by
the tree), and then the child was asked to describe the second picture, which
showed dierent characters performing dierent actions. Children aged 4 and
above showed evidence of priming (they were more likely to produce a passive
sentence when describing the second picture if the had just heard this struc-
ture), but younger children did not. However, there was priming even in the
younger children in another condition, when there was lexical overlap between
the two sentences that is to say, when the experimenters sentence contained
lexical material that the child could use in her own sentence (e.g. He got
VERBed by it).
Thus, experimental research conrms that early grammatical knowledge is
lexically specic. It also shows that the emergence of general word-order
schemas is a very gradual process. The preferential looking studies suggest
that some form of verb-general knowledge of the transitive construction
begins to develop about age 2;6, but it takes another year or so before chil-
drens representations are strong enough to support production. Verb-general
knowledge about less frequent structures such as the passives takes even
longer to develop.
2.7 Mosaic acquisition

Even if early multi-word utterances are constructed using formulaic frames of
the kind described in the preceding subsection, such frames can only account
for certain aspects of language development. Acquiring the full adult system
requires much more than learning the positional requirements of individual
lexical items. The child will eventually be able to add various inections to
thousands of stems and command a wide range of complex syntactic con-
structions. Underlying these, as every linguistics undergraduate knows, are
some rather abstract rules, which she must learn at some stage. Conventional
linguistic wisdom has it that children are able to infer these rules with ease and
after relatively little exposure. Moreover, once a rule is learned, it is immedi-
ately applied in all the contexts that require it. It is striking, Chomsky tells us,
that advances are generally across the board (1964: 39; see also Chomsky
1999).
Or are they? This doesnt appear to be corroborated by longitudinal
research, which suggests that grammatical development is typically slow and
piecemeal. For example, research on the acquisition of grammatical mor-
phemes shows that children initially learn to add the present-participle and
past-tense endings to one verb at a time and that it can take quite a long time
for the endings to penetrate the entire linguistic system (Brown 1973; Cazden
1968):
A considerable period of time elapses between the rst appearances of

a morpheme and the point where it is almost always supplied where
required. The progressive ending, for example, is rst supplied 100

percent of the time in obligatory contexts in sample 16. This comes 16
months later than sample 1 when -ing was supplied 50 percent of the
time. We can be sure that there were some progressives and some plurals
in Sarahs speech before we began to transcribe it, so that the time elaps-
ing between rst occurrences and a perfect performance is even
longer . . . It is true of all the grammatical morphemes in all three chil-
dren that performance does not abruptly pass from total absence to reli-
able presence. There is always a considerable period, varying in length
with the particular morpheme, in which production-where-required is
probabilistic. This is a fact that does not accord well with the notion that
the acquisition of grammar is a matter of the acquisition of rules, since
the rules in a generative grammar either apply or do not apply. One
would expect rule acquisition to be sudden. (Brown 1973: 298)
Later research has tended to corroborate Browns suggestion that the acqui-
sition of grammatical inections is a process of piecemeal addition rather than
sudden revelation. Bloom et al. (1980) reported that the irregular-past,
present-participle and third-person-singular inections tend to be used with
dierent populations of verbs: -ing predominantly with durative non-comple-
tive verbs such as eat, play and write; the irregular past with non-durative com-
pletive verbs such as nd, get and come; and -s predominantly with go and t
(as in This goes/ts in here). More recently, similar ndings have been reported
by Tomasello (1992: see above), Pizzuto and Caselli (1992), Caselli et al. (1993)
and Rubino and Pine (1998).
Rubino and Pines work is particularly interesting since it shows that verb-
specic learning can occur even at a time when the child already has general
rules. They studied the acquisition of subjectverb agreement by a 3-year-old
child learning Brazilian Portuguese. By this stage, the child appeared to have
mastered the agreement system of his language, since the overwhelming
majority of his verb forms had the correct inections: the error rate was only
3 per cent. However, Rubino and Pine were able to show that this low overall
rate was due to very good performance on a few very frequent categories (spe-
cically, the singular inections) and masked much higher error rates in other
parts of the system. Plural forms accounted for only 3.4 per cent of the verb
forms used by the child (which suggests that he avoided using the plural), and
had much higher error rates: 23 per cent for the rst person plural a gente and
43 per cent for the third person plural. (Second person plural forms did not
occur at all in their corpus of almost 1,500 utterances.) Moreover, seven out
of the twelve utterances exhibiting correct 3PL agreement contained one of
two very frequent verb forms. Thus, there is little evidence for the mastery of
plural inections: the child appeared to be learning the system piecemeal.
But perhaps the most convincing evidence for such mosaic acquisition
comes from studies of the development of rules for auxiliary verb placement
in English. Early studies of the development of interrogatives and negatives

(Bellugi 1971; Klima and Bellugi 1966) reported that progress is across the
board: that is to say, auxiliaries appear in all the syntactic contexts that require
them at more or less the same time; and, moreover, once the child learns the
rule of subjectauxiliary inversion, she applies it to all the auxiliaries she
knows. However, these early studies were more concerned with writing gram-
mars for early stages of language development than with documenting tran-
sitions from one stage to the next, and so simultaneous development was
asserted rather than demonstrated. More recent research on auxiliaries shows
a very dierent picture.
The rst chink in the argument for across-the-board acquisition of English
auxiliaries was an observation made by Bellugi and Klima themselves that
subjectauxiliary inversion seemed to appear in yes/no questions rst and
spread to WH questions only later in development (Bellugi 1971; Klima and
Bellugi 1966). However, Bellugi had a good story to tell about why this should
happen: children at this stage, she argued, know both the subjectauxiliary
inversion and the WH fronting rule, but due to limited processing power are
unable to apply both rules in a single derivation. Hence, we get correctly placed
auxiliaries in yes/no questions, which require inversion only, and uninverted
auxiliaries in WH questions, which require WH fronting as well as inversion.
Bellugis explanation, though elegant, is far from water-tight, as she herself
acknowledges. If limited processing resources prevent children from perform-
ing both transformations at the same time, they should produce not only ques-
tions with WH fronting but no inversion (Where you will go?), but also inverted
questions without WH fronting (Will you go where?). Such structures,
however, do not occur.
Moreover, new evidence soon began to accumulate that the acquisition of
the rules for auxiliary placement in questions is disturbingly piecemeal. In
their study of auxiliary placement in yes/no questions, Kuczaj and Maratsos
(1983) found that children preposed some auxiliaries but not others. The study
was prompted by the observation that Kuczajs son Abe rst began using
would and could in declaratives at 2;9 and 2;11 respectively, but did not use
them in questions until 3;5, although he did produce interrogatives with other
modals, including will and can. Kuczaj and Maratsos then looked at a larger
cross-sectional sample of 14 children between the ages of 2;6 and 5;6. They
found the same pattern for could in the speech of seven children and for would
in the speech of three children. They also noted that none of the children over-
generalised the inversion rule to auxiliary-like predicates such as better, gotta,
hafta and wanna, which also suggested that they did not have a general rule
involving auxiliaries as a class.
Similar ndings from a study that used experimental methods with older
children (58-year olds) were reported by Major (1974). Major asked children
to imitate sentences with modal verbs, to transform declaratives into negatives
and interrogatives, and to add tags to model sentences. Her general conclusion
was that although all of the children were familiar with some of the uses of all
modals and semi-modals, knowledge about the constructions in which a par-
ticular word could occur was not generalised from one member of the class to
another. The children generally performed well on high-frequency modals
(will, would, can, could, should) but error rates on may, might, shall and must
were very high: for example, they were 2560 per cent correct on negatives with
may, depending on age group, and 020 per cent correct on questions with the
same verb. When faced with the task of transforming You may go home into a
question, most of the children (20) simply substituted a more familiar modal
with a similar meaning that is to say, they produced well-formed questions
with will, would, can or could. Six children produced the target structure (May
you go home?); one simply repeated the stimulus sentence using a rising into-
nation. Most revealing, however, were the ungrammatical sentences produced
by some of the children. Two children responded with Do you may go? and 11
added another modal at the beginning of the sentence, producing Can you may
go? Could you may go?, Will you may go?, or Would you may go? Dierences
between the modals showed up even on the imitation task. For example, 83 per
cent of the 5-year-olds imitated He should have nished his breakfast and We
could have gone to the show, but only 8 per cent were able to reproduce faith-
fully an analogous sentence with the auxiliary can (Sally can have chased the
dog). Even more patchy was the childrens command of tag questions. Every
child produced at least one acceptable tag, but error rates were very high, and
no child produced tags with may, might, shall or ought to. Moreover, some
added the same tag to nearly all sentences, which also suggests that they had
not fully mastered the structure.
Modals in WH questions show a similar pattern of development. An elic-
ited imitation experiment conducted by Kuczaj and Brannick (1979) showed
that children learn to place modals correctly rst after one or two WH words
(usually what and where), then after another word, and so on: in other words,
they learn the rule piecemeal. Similar results have been reported by Labov and
Labov (1978) and Ingram and Tyack (1979).
However, Kuczaj and Brannick also found that the children did eventually
generalise, or at least some did. Fifteen of their subjects (out of 60) overgen-
eralised the rule to questions with how come, which resulted in ungrammati-
cal utterances such as *How come will the wolf howl at the moon? However, six
of the children who overgeneralised failed to use modals at all with how long;
and ten correctly repeated questions with how come + past verb (How come the
turtle fell on the ground?) in other words, the overgeneralised rule was not
fully general!
There is also evidence suggesting that the knowledge that children acquire
is not only lexically specic but also construction-specic: that is to say, chil-
dren learn about specic words in specic constructions. Rowland and Pines
(2000) detailed analysis of WH questions produced by a single child revealed
that he tended to prepose some auxiliaries after some WH words and use
declarative word order with other WH+AUX combinations, and that the com-
binations that were preposed were more frequent in the input than the non-
preposed ones. Fletcher et al. (2003), using an elicited production task, showed
that 3-year-olds performed very well on WH questions with the copula (which
are very frequent in the input), but were unable to produce structurally iden-
tical questions with the auxiliary should. However, when the specic structure
(i.e. WH questions with should) was modelled for the children in the course of
the experiment, performance improved dramatically. In another experiment,
the children were trained with inverted Y/N questions with is, can or should
and tested on WH questions with the same auxiliary. In this case, the training
did not improve performance: in other words, the children were not able to
generalise knowledge about auxiliary placement in Y/N questions to WH
questions.
Another telling example of piecemeal learning is provided by Todds (1982)
study of the acquisition of tag questions by his daughter Audrey. Audrey rst
began to use tags at 3;2, but except for two frozen forms which occurred very
early on, all her tags were armative. This resulted in ungrammatical utter-
ances when the tagged sentence was also armative: I look funny, do I?, Were
going home, are we? (both said with a rising intonation).5 She began to use neg-
ative tags productively at 4;2, but only with two verb forms, is and will. At 4;11,
she added wasnt NP? to her repertoire of negative tags. Finally, about 5;3,
negation spread to all auxiliaries used in tags and was overgeneralised to neg-
ative sentences, where she had previously supplied the correct form (*There
isnt very much people here, isnt there?).
Taken together, the studies reviewed here provide very strong evidence that
the development of auxiliary placement rules in English is a step-by-step,
bottom-up process involving the gradual accumulation of specic bits of
knowledge which are only later integrated into a system. Children do not
acquire an abstract rule of subjectauxiliary inversion: they learn about spe-
cic forms in specic constructions. They may subsequently develop broader
rules by generalising over these low-level patterns, but the early rules are highly
specic.
This conclusion is corroborated by a more comprehensive study of the
English auxiliary system by Richards (1990). Richardss monograph surveys
the development of auxiliary verbs in seven children over a period of some
eight months. Using recordings of spontaneous speech as well as an elicited
imitation task, he charts the emergence of all allomorphs of the core auxiliary
verbs in all syntactic contexts in which the auxiliary occurs: declaratives, neg-
atives, Y/N questions, WH questions, and emphatic and elliptical uses, includ-
ing tags. The study revealed wide-ranging individual dierences and a great
deal of piecemeal learning:
Across the board developments did not occur, though there were
periods of rapid development for some children. Instead, progress was
typied by piecemeal additions to the range of forms and the range of

linguistic contexts and syntactic frames in which they were used.
(Richards 1990: 214)
3. A case study: questions
We saw in the preceding section that childrens early use is dominated by xed
phrases and low-scope patterns. There is considerable disagreement, however,
about the relationship between these early formulas and the more general pat-
terns which emerge later. Some researchers (e.g. Bates et al. 1988; Brown and
Hanlon 1970; Krashen and Scarcella 1978) have argued that formulas are a
developmental dead-end, or at best a side street, a kind of linguistic crutch
which allows children to get by in the early stages of development, but which
is eventually superseded by real syntax with standard combinatorial rules.
Others (e.g. Lieven et al. 1992; Pine and Lieven 1993; Tomasello 2000, 2003)
have suggested that rote-learned phrases play a critical role in the development
of productive use. On this account, the transition to adult syntax is quite
gradual and does not involve qualitative shifts.
In this section, I will address this issue by examining in some detail the
acquisition of a particular construction, or rather family of constructions,
namely, English questions. Question constructions can provide more conclu-
sive evidence about the role of lexically specic learning in development than
argument structure constructions. It makes sense for children to be conserva-
tive when acquiring the latter, since dierent verbs have dierent subcategor-
isation requirements. Questions, on the other hand, share some very general
structural properties (often captured by means of two abstract rules: WH
fronting and subjectauxiliary inversion), so if children acquire fully general
knowledge from the start, this is where we would expect to nd evidence of it.
The acquisition of question constructions is also interesting from a theoreti-
cal point of view, since the two movement rules have played a very important
role in the development of the generative linguistics.
I have already noted that children sometimes acquire interrogative formu-
las such as Whats that? and Wheres X? in early stages of development, long
before there is any evidence of true syntax. We will now examine childrens
questions more systematically in order to determine how prevalent such for-
mulaic uses are, how long they persist, and whether there is any evidence of
developmental discontinuities in the transition to more productive usage
which might signal the acquisition of rules.
3.1 Early interrogative utterances

Lieven (2003) studied syntactic questions (multi-word utterances containing a
preposed auxiliary and/or a preposed WH word) in the speech of four 2-year-
olds recorded for ve hours a week over a six-week period. This recording
regime produced samples which are ve to ten times denser than those used in
most studies and comprise 7 per cent to 10 per cent of each childs linguistic
experience during this period. With such relatively large samples, it is possible
to identify recurrent word combinations which are likely to be retrieved from
memory as ready-made units.
Lieven was interested in two types of preconstructed units, or formulas: lex-
ically xed phrases (e.g. Whats that?) and frames with a slot which can be
lled by a variety of expressions (e.g. Wheres X?). For the purposes of the
study, Lieven assumed that an utterance was formulaic if the corpus contained
at least three tokens of the same combination of words produced on three
dierent occasions. A frequency of three or more might not seem like very
much; however, extrapolating from the sample, we can estimate that the child
is likely to have produced 3040 tokens of the utterance in question during the
six-week period covered by the study, so the assumption does not seem unrea-
sonable.6
Lievens ndings are presented in Table 9.1. As we can see, the children had
from 6 to 19 dierent interrogative formulas, and these accounted for between
78 per cent and 98 per cent of their syntactic questions (mean 92 per cent).
The table also contains the results of an analysis of a comparable corpus of
data from a Polish-speaking child, Marysia (Dabrowska and Lieven 2003).
The gures are very similar, thus providing some cross-linguistic validation for
Lievens ndings.
Table 9.1 Formulaic questions in the speech of five 2-year-olds

Brian Chris Annie Delia Marysia
MLU (words) 1.72 1.98 2.12 2.23 1.87

Number of formulas 6 17 11 19 18
Total syntactic questions 346 434 409 300 695
Of which:
Fixed phrase 70% 45% 17% 17% 80%
Frame with slot 28% 49% 77% 61% 14%
Non-formulaic 2% 6% 5% 21% 6%
Note: The figures for the English-speaking children (Brian, Chris, Annie and Delia)
are based on Lieven (2003). The figures for the Polish-speaking child, Marysia, are
based on Dabrowska and Lieven (2003).
The data for the four English-speaking children is arranged from least
advanced to most advanced, as measured by MLU. This reveals an interesting
pattern: the longer the MLU, the lower the proportion of xed formulas and
the higher the proportion of frames with slots. The most advanced child,
Delia, also has a relatively high proportion of non-formulaic utterances. This
could mean that Delia had acquired some more general knowledge about
interrogative utterances. However, Lieven notes that all but two of Delias
non-formulaic questions could be traced back to utterances in the input in a
relatively straightforward way: that is to say, they were either exact repetitions
of things the mother had repeatedly said before, or were based on frames
which could have been derived from the input (for example, the utterance
Whats Mama making? may be based on a Whats Mama V-ing? frame, derived
from parental utterances such as Whats Mama doing?, Whats Mama
cooking? etc.).
3.2 From formula to schema

Lievens study conrms that early questions are highly formulaic. Further-
more, the fact that the least advanced child used primarily invariant formulas,
while more advanced children relied more on formulaic frames and also pro-
duced a higher proportion of non-formulaic utterances, suggests that devel-
opment at this stage may involve a progression from invariant formulas to
more schematic templates. However, it is also possible that the more formulaic
children were relying on more holistic learning strategies, which, according to
some researchers (e.g. Bates et al. 1988), may be less eective. Thus, to deter-
mine whether development indeed proceeds from formula to schema, we need
to look at longitudinal data.
In this section we examine the development of questions in three children
from the CHILDES database (MacWhinney 1995): Naomi, Sarah and Shem
(Table 9.2). Because the corpora for these children are larger (though less
dense) than those used by Lieven, I will adopt a slightly more stringent criter-
ion for formulaicity: a question type will be considered formulaic if there are
at least ve instances of it in the childs speech. As in Lievens study, the term
formula will subsume frames with a single slot as well as lexically xed
phrases.
Figure 9.1 shows the proportion of the two types of formulas as well as non-
formulaic questions in the speech of the three children during the period
covered by the study. As we can see from the graphs, virtually all of the ques-
tions produced by the 2-year-olds are formulaic by our criterion: these nd-
ings thus replicate the results of Lievens study. As the childrens linguistic
systems develop we see a steady decrease in the proportion of xed formulas
Table 9.2 Data for the longitudinal study

Corpus Corpus size Number of
(child words) Age range syntactic questions
Naomi (Sachs 1983) 38,729 1;83;8 1,439

Shem (Clark 1982) 65,651 2;23;2 1,173
Sarah (Brown 1973) 89,086 2;35;1 2,006
Figure 9.1 Formulaic and non-formulaic questions in the three children.

Note: The ages on the Y axis represent the midpoint of each period.
and a corresponding increase in the proportion of more creative utterances.

This pattern is most clearly visible in Naomi and Sarah, where we have data
spanning longer periods, but it is also discernible in Shem.
Several further observations are in order. First, each child acquired a dier-
ent set of question formulas (cf. Table 9.3), although, not surprisingly, some
Table 9.3 The childrens 10 earliest formulas (in order of emergence)

Naomi (1;81;11) Shem (2;2) Sarah (2;32;8)
MLU 1.42.6 MLU 2.8 MLU 1.62.4
whats this/that? whats this/that? where NP?

what do you VP? (VP = see) wheres NP? whats this/that?
wheres NP? what are (you) doing? are you AP? (AP = bad)
who VP? (VP = did that) what (is) NP doing? who V it/that?
where is NP? (NP = it) whats LOC? where S? (S = gone)
where did it go? whats NP? where NP go?
is this NP? what is that? what VP? (VP = happen)
is it AP? what happened to NP? where is it?
wheres NP go? (NP = it) what S? why S? (S = need them)
is NP going to work? where NP go? who(s) this?
Note: The figures in the column headings indicate the time span during which the
first 10 formulas emerged and the childs MLU at the beginning and the end of this
period. If there was no variation in a particular slot during this early period, the only
attested filler is given in parentheses after the formula. The very first recording in the
Shem corpus already contains 10 formulas, suggesting that they developed earlier.
(e.g. whats this/that?, wheres NP?) show up in all three. Secondly, there are
marked dierences in the rate of development. Naomi acquired her tenth
question formula at age 1;11, while Sarah needed another nine months to
reach the same milestone. Finally, and most importantly, formulaic utterances
also feature very prominently in later development: even at age 3 and 4, a sub-
stantial majority of the childrens utterances can still be accounted for in terms
of a relatively small number of lexically specic units.
Thus the longitudinal data appears to support the scenario proposed earlier,
namely, that development proceeds from xed phrases through formulaic
frames to more abstract patterns. Furthermore, there are no abrupt changes
in behaviour which might signal the acquisition of a general rule. What we see
instead is a slow but steady increase in the proportion of creative utterances,
suggesting a series of small changes. To understand what such small changes
might involve, it will be helpful to examine the development of a few formu-
las in more detail.
3.2.1 Sarahs who VP?

Sarahs earliest who questions were who did it? (produced at 2;4.26) and who
did that? (produced two weeks later). During the next four weeks, she pro-
duced another six tokens of this formula, and a further 10 were attested in
later recordings. (Note that it, that and this are frequently substituted for each
other, but not for other forms, in a variety of constructions. Because the vari-
ation is very limited, it is not regarded as dening a slot, so who did it/that? was
treated as a single xed formula.) A related lexically xed phrase appeared at

2;6;20:
(6) a. Who broke it? (2;6.20)

b. Who broke that? (2;7.5)
The corpus contains seven tokens of this formula in total. Two months later,
Sarah began using other verbs in this construction:
(7) a. Who rip that? (2;8.25)

b. Who bought that? (2;8.25)
c. Who done this? (3;0.18)
d. Who drop this? (3;3.20)
e. Who lost it? (3;6.6)
Meanwhile, four other formulas with who appeared in Sarahs speech: who
this?, who is it?, whos that? (all attested for the rst time between 2;8 and 2;11;
note that dierent pronouns were used with dierent forms of the copula),
and, a little later, who(s) goin(g to) VP?, rst attested at age 3;6. Finally,
almost a year and a half after the emergence of the earliest formula, we get
subject questions with who in sentences containing non-pronominal objects
and non-transitive verbs, e.g.:
(8) a. Who didnt work? (3;9.26)

b. Who said cant? (3;10.16)
c. Who put the bandaid on? (4;1.11)
d. Who scribbled on it? (4;3.19)
Thus, development seems to have begun with two xed formulas, who did
it/that? and who broke it/that? The formulaic frame who TRANS.V it/that?
which emerged somewhat later seems to be a generalisation over the two lexi-
cally xed phrases (and possibly others which have not been captured in the
sample). The fully general template for producing questions with who in the
subject slot (who VP?) appeared considerably later, after Sarah had mastered
a number of more specic who-subject formulas.
3.2.2 Naomis whats NP doing?

Naomis rst recorded attempt at such questions was a telegraphic What
doing? produced at 1;11.11. Ten days later she imitated a complete question
with both a subject and a contracted auxiliary, Whats Mommy doing? The
transcripts for the next four weeks contain nine further occurrences of this
utterance in exactly the same form. Then, at 2;0.18, she began producing
similar questions with subjects other than Mommy:
(9) a. Whats donkey doing? (2;0.18)

b. Whats toy doing? (2;0.18)
c. Whats Nomi doing? (2;0.18)
After this, there are many questions with the names of various people and toys
in the subject slot. A week later, Naomi began substituting other verbs for
doing:
(10) a. Whats Mommy holding? (2;0.26)

b. Whats Georgie saying? (2;1.9)
At this point, she could be said to have acquired a template with two slots
(Whats NP TRANS.V-ing?). Finally, just before her third birthday, she began
to use the uncontracted form of the auxiliary, which suggests that she had ana-
lysed whats into its component morphemes:
(11) a. What is the boy making? (2;11.17)

b. What is Andy doing? (2;11.18)
Thus, we see a similar progression from a rote-learned lexical chunk (Whats

Mommy doing?) through increasingly general schemas (Whats NP doing?,
Whats NP TRANS.V-ing?) to a fairly general template (What BE NP
TRANS.V-ing?) which can be used to produce a wide variety of interrogative
utterances.
3.2.3 Naomis can I VP? and could I VP?

Questions with the modal auxiliary can appeared in Naomis speech shortly
before her second birthday.
(12) a. Can I get down? (1;11.9)

b. Can I get up? (1;11.9)
c. Can lie down? (1;11.16)
These early questions were quite stereotypical: the auxiliary was always placed
at the beginning of the sentence, and the subject, if present, was always the
rst person pronoun I; when the subject was left out, it is clear from the
context that the agent of the action was Naomi herself. This suggests that at
this stage, she did not have a general auxiliary placement rule, but merely a
formula for requesting permission.
With time Naomis usage gradually became more exible. First she acquired
a variant of the permission-seeking formula with the auxiliary could instead
of can. Like the canonical form of the formula, this variant was produced
sometimes with and sometimes without an explicit subject.
(13) a. Could do this? (1;11.21)

b. Could eat that? (1;11.21)
c. Could I throw that? (2;0.3)
Then, at 2;0.28 seven weeks after the rst recorded use of the formula we
nd the rst question with a subject other than I.
(14) a. Can you draw eyes? (2;0.28)

b. Please can we do that? (2;8.14)
So, as in the previous two examples, early usage was highly stereotypical and
gradually became more exible as new slots opened up inside the formula.
3.3 Where do non-formulaic utterances come from?

We have seen that questions produced by 2-year-olds are extremely stereotyp-
ical, in that virtually all are instances of one of a relatively small number of
formulas. Older children continue to use such stereotypical phrases, but also
produce increasingly high proportions of apparently more creative utterances.
Where do these non-formulaic questions come from, and what do they tell us
about the childrens grammatical abilities?
It is important to realise that many of the utterances which were classied
as non-formulaic according to the frequency criterion (ve or more child
tokens in the corpus) could actually be instances of well-practised routines.
The transcripts analysed here contain a tiny sample of the childrens linguis-
tic experience somewhere between 1 per cent and 2 per cent of what they said
and heard so even utterances that the child had produced hundreds of times
may only occur once in the sample, or not be found at all. Furthermore, 3050
per cent of the non-formulaic utterances, though rare in the child, occur ve
or more times in the input, so there is a good chance that they have been mem-
orised. (Note that the mean frequency of utterances which occur ve times in
a 1 per cent sample of the input is 500, assuming the sample to be representa-
tive.)
On the other hand, some non-formulaic questions are without doubt crea-
tive. This is evident from the fact that they contain relatively high proportions
of errors of commission (cf. Table 9.4).7 What is more, in contrast to the errors
found in formulaic utterances (see below), these tend to be quite unsystematic,
which suggests that they are one-o innovations produced when the child
attempted to go beyond the well-practised formulas and say something he or
she had not yet learned to say.
The most interesting subcategory of non-formulaic utterances is those
which are grammatical yet cannot be regarded as delayed imitations of
phrases which occur frequently in the input. Such utterances do indicate that
the child is moving beyond lexically specic formulas. However, as we saw in
Table 9.4 Errors of commission in formulaic and non-formulaic questions

Number of questions % errors of commission
Child Formulaic Non-formulaic Formulaic Non-formulaic
Naomi 1,271 168 2 12

Shem 1,058 115 14 27
Sarah 1,681 325 7 16
the preceding subsection, they dont necessarily require abstract movement

rules, since they could be produced by substituting new material into more
abstract templates with two or even three slots. For example, can we do that?
could be constructed by inserting we and do that into the NP and VP slots in
can NP VP?; whats Georgie saying? could be constructed by inserting Georgie
and say into the appropriate slots in whats NP TRANS.V-ing?; and so on.
3.4 Evidence for piecemeal learning

On the other hand, the fact that the childrens questions can be described in
terms of lexically specic formulas and templates does not entail that they
should be so described, since the same data could also be accounted for (more
economically) in terms of general rules. In order to determine how abstract or
concrete childrens syntactic representations really are, we need to turn to
other sources of information. As indicated earlier, there are several experimen-
tal studies suggesting that knowledge about particular auxiliaries and WH
words is indeed acquired piecemeal. There are also some interesting patterns
in the longitudinal data examined here which support this interpretation (see
also Dabrowska 2000 for a more detailed discussion of the Naomi data).
3.4.1 Auxiliary placement

Y/N questions in English are normally marked by a preposed auxiliary (Can
I do that?), with rising intonation acting as a secondary marker. Another
option is to use only rising intonation (I can do that?). Young children often
rely on this simpler method; then, as they acquire the adult system, the pro-
portion of inverted questions (i.e. questions with a preposed auxiliary)
increases (see Figure 9.2).
It is noteworthy that the course of development is dierent in dierent chil-
dren. In Shem, we see a steady increase in the proportion of inverted question
during the entire period of observation. Sarah shows a similar pattern, except
that progress levels o after age 4;0. Naomis development, in contrast, follows
a very clear U-shaped curve. In the early stages (before age 2), the number of
questions with preposed auxiliaries was relatively high, but they were very
stereotypical. As she became more adventurous and started asking questions
about a wider variety of situations, the proportion of inverted questions
Figure 9.2 Proportion of inverted Y/N questions

dropped. Then, from about 2;6 onwards, it started rising again until, towards
the end of the period covered by this study, it reached levels similar to those
found in the input.8
The overall inversion rates in Figure 9.2 mask some interesting low-level
patterns suggesting that the children acquired knowledge about specic com-
binations of words rather than general rules. To show such eects, one would
ideally want to compare the proportions of inverted and uninverted questions
for each auxiliary + subject combination at dierent stages of development.
However, such an analysis would require a corpus containing a much larger
sample of each childs speech for each developmental period. Fortunately,
some patterns are discernible even if we collapse gures across auxiliaries or
subjects; and it is also possible to conduct a more ne-grained analysis of
questions with the most frequent auxiliary, namely do.
Table 9.5 shows the proportions of inverted yes/no questions with the most
frequent auxiliaries in the three children, and, for comparison, in one of the
mothers.9 To avoid collapsing across very long periods, and also to give some
sense of development, the Sarah corpus has been split into two parts: Sarah1
covers the period from 2;3 to 3;7, and Sarah2, from 3;8 to 5;1.
Table 9.5 Inverted Y/N questions according to auxiliary

Naomis
Naomi Shem Sarah1 Sarah2 input
Auxiliary N % N % N % N % N %
can 32 100 36 80 21 88 138 91 68 97

will 6 86 3 10 78 38 83 0
would 11 100 1 0 3 40 100
aux-bea 21 54 12 43 2 40 68 93 72
copulaa 78 76 34 44 22 59 227 75 90 68
do/does 36 39 55 37 4 3 169 40 207 62
did 8 53 5 31 10 67 88 77 93 72
Note: Percentages are provided only when there were at least eight questions with a
given auxiliary (inverted and uninverted).
a
Present-tense forms only
It is clear from the gures in the table that the children preposed some aux-
iliaries considerably more often than others. All three used the inverted form
by far the most frequently with can. This is also true of the mother, so the chil-
drens behaviour may be simply a direct reection of the input: they heard
both inverted and uninverted questions with the other auxiliaries, so they pro-
duced both types; with can, they virtually never heard uninverted questions,
so they rarely produced them themselves. The remaining dierences are partly
attributable to order of acquisition (since the data was collected over a period
of time, if a child acquired the inverted pattern for a particular auxiliary fairly
late, the overall inversion rate for that auxiliary will be low) and partly to the
fact that dierent auxiliaries follow dierent developmental trajectories (for
example, although questions with do and did appeared in Sarahs speech at
about the same time, she used the latter much more consistently than the
former).
Such dierences between inversion rates for individual auxiliaries can be
accommodated in a rules-based approach: even if children acquire a general
rule, they still need to learn about the lexical properties of each auxiliary, and
it is perfectly possible that they acquire such knowledge at dierent times for
dierent auxiliaries. Much more problematic for such approaches is the fact
that there are also substantial dierences in the frequency of inversion with
dierent subjects.
Consider rst questions with do. Do is used when there is no other auxiliary
in the sentence, so the uninverted (i.e. intonation-only) analogues of such
questions normally lack an auxiliary (You want some?), except when they are
emphatic (You do want some?). Table 9.6 shows frequency of inversion with
Table 9.6 Questions with present tense forms of the auxiliary do

Naomis
Naomi Shem Sarah2 input
Person and
number N % N % N % N %
1sg 0 0 0 0 21 44 0
2sg 29 52 53 54 92 35 153 61
3sg 7 23 1 5 45 52 43 67
Plural 0 1 7 11 48 11 55
36 38 55 38 169 40 207 62
given subject (inverted and uninverted).
rst, second and third person singular and with plural subjects in the three
children and, for comparison, in one of the mothers.10 (The gures for plural
subjects have been collapsed to allow meaningful comparisons, as such ques-
tions are relatively rare.) As we can see, the mothers inversion rates are very
similar for all person/number combinations. The children, on the other hand,
put do before some subjects considerably more often than before others, which
is precisely what one would expect if they initially acquired knowledge about
particular combinations of subjects and auxiliaries.11
Moreover, there are dierences between the children in this regard. Naomi
and Shem invert by far the most frequently with second person subjects in
fact, Shem almost never uses do with any subject other than you. The most
plausible explanation for this early prociency with second person subjects is
frequency in the input: children hear a great many questions of the form do
you VP?, so they have plenty of opportunity to pick up the pattern. Questions
with third person subjects are also quite frequent, but considerably more
varied (does it VP?, does he VP?, does she VP?, as well as some with lexical sub-
jects), so there are fewer opportunities to learn any particular pattern. Do
questions with rst person and plural subjects, on the other hand, are rare in
the input; and it is also these combinations that Naomi and Shem have most
diculty with.
Sarah, however, shows a very dierent pattern: she is least likely to invert
with second person subjects. It seems that several factors have conspired to
produce this eect. First, as is evident from Figures 9.1 and 9.2, Sarahs early
development was quite slow compared to the other children. She appeared to
have particular problems with the auxiliary do (cf. Table 9.5), possibly because
it was often contracted to [d] or [d ] in the input (as in dyou want an apple?),
and hence not very salient acoustically. Consequently, by the time she learned
her rst do-frames, she already had several well-entrenched intonation-only
formulas such as you want NP? and you know what? which often pre-empted
the emerging templates with auxiliaries. We see this very clearly when we
compare the frequencies of lexical verbs in inverted and uninverted questions.
The latter are dominated by high-frequency verbs: 52 per cent have either want
or know as the main verb, and 73 per cent contain one of just six verbs; the
corresponding gures for inverted questions are 29 per cent and 56 per cent
respectively. Thus Sarahs development, although it followed a dierent course
from that of the other two children, also shows evidence of lexically specic
learning.
Similar dierences in the proportion of inverted and uninverted questions,
depending on the subject, are found with other auxiliaries (cf. Table 9.7); and
Table 9.7 Inverted Y/N questions according to person and number in

questions with auxiliaries other than do
Naomis
Naomi Shem Sarah1 Sarah2 input
Auxiliary N % N % N % N % N %
1sg 28 85 31 74 25 81 129 87 20 65
2sg 32 67 20 53 23 61 177 80 268 79
3sg 69 45 33 29 22 43 298 75 176 76
Plural 7 54 4 3 22 50 20 69
Total 136 55 88 44 73 59 626 77 484 77
given subject (inverted and uninverted).
again, the dierences are small and not statistically signicant in the mother,
and highly signicant in the children.12 It is interesting to note that the rela-
tive proportions of inverted and uninverted questions with dierent subjects
are dierent from with do: all three children invert most often in questions
with I, and least often with third person singular and plural subjects.
The pattern of inversion/uninversion with the pronoun I is particularly
interesting, and merits a more detailed discussion. Questions with rst person
subjects are relatively rare in the input. Only 3 per cent of the Y/N questions
addressed to Naomi had rst person subjects, and almost half of these had
the form can I VP? Children, on the other hand, ask such questions quite
often, thus oering the analyst an excellent opportunity to determine to what
extent they have generalised the patterns in the speech addressed to them. If
inversion rates with I are as high as with other subjects, this would suggest that
children have fairly general rules; and conversely, low inversion rates would
entail failure to generalise. As we saw earlier, children rarely produce syntac-
tic questions with rst person subjects and the auxiliary do; but inversion rates
for I with other auxiliaries are quite high (cf. Table 9.7). However, a closer
analysis reveals some striking patterns. The high inversion rates that we see in
Table 9.7 are due to the fact that the vast majority of the childrens questions
with rst person subjects are can I VP? questions the only frequent rst
person pattern in the input which all three children consistently invert (see
Table 9.8). If we exclude questions with can, inversion rates with rst person
subjects will be very low, except for the second half of the Sarah corpus, which
covers data for a relatively late period (3;85;1). Moreover, if we scrutinise the
constructions in Table 9.8 more closely, we will see that there is very little
overlap in the inverted and uninverted combinations of rst person subject
and auxiliary: that is to say, for any particular combination, each child either
consistently inverts (e.g. can I . . . ?) or consistently uses declarative word order
(Im . . . ?). (The only exception to this pattern is, again, the second part of the
Sarah corpus: but it is not surprising to nd evidence of more general knowl-
edge in a child of this age.) Thus, the pattern of inversion/uninversion with I
provides particularly clear evidence for piecemeal learning: children acquire
knowledge about particular subjectauxiliary combinations rather than
general rules which apply across the board.
3.4.2 Contraction
In addition to their canonical or full forms, the English copula and some aux-
iliaries have contracted forms (e.g. s, re) which cliticise to the preceding word.
The contracted forms predominate in informal adult speech, and are also
quite frequent in childrens earliest questions. Many children, however, go
through a stage when they prefer the full forms, sometimes after a period of
fairly consistent use of contracted forms. This marks a departure from the
adult pattern which is usually interpreted as indicating that the child has seg-
mented previously unanalysed forms (e.g. whats, wheres) into their component
Table 9.8 Inverted and uninverted questions with first person singular subjects
Corpus Inverted questions Tokens Uninverted questions Tokens
Naomi can I . . . ? 25
could I . . . ? 3
I/me V-PRES . . . ? 8
Im . . . ? (copula) 1
Im . . . ? (auxiliary) 1
I cant . . . ? 1
Shem can I . . . ? 27
could I . . . ? 2
is I . . . ? 1
should I . . . ? 1
I V-PAST . . . ? 1
I V-PRES . . . ? 12
Im . . . ? (auxiliary) 3
Ill . . . ? 1
I cant . . . ? 1
I was . . . ? (copula) 1
I was . . . ? (auxiliary) 1
Sarah1 will I . . . ? 2
shall I . . . ? 1
can I . . . ? 18 I can . . . ? 1
did I . . . ? 4 I V-PAST . . . ? 2
do I . . . ? 1 I V-PRES . . . ? 33
I cant . . . ? 1
I didnt . . . ? 1
I be ..? (copula) 1
Sarah2 could I . . . ? 11
shall I . . . ? 1
am I . . . ? (copula) 7
was I . . . ? (copula) 1
can I . . . ? 77 I can . . . ? 2
should I . . . ? 5 I should . . . ? 1
did I . . . ? 21 I V-PAST . . . ? 6
am I . . . ? (auxiliary) 5 Im . . . ? (auxiliary) 4
do I . . . ? 21 I V-PRES . . . ? 27
dont I . . . ? 1 I dont . . . ? 1
I was . . . ? (auxiliary) 1
I cant . . . ? 1
I didnt . . . ? 1
I will . . . ? 1
Naomis didnt I . . . ? 3
mother shall I . . . ? 1
did I . . . ? 1
would I . . . ? 1
can I . . . ? 9 I can . . . ? 1
am I . . . ? (copula) 1 I am . . . ? (copula) 3
am I . . . ? (auxiliary) 1 I am . . . ? (auxiliary) 1
should I . . . ? 3 I should . . . ? 3
I shouldnt . . . ? 1
I cant . . . ? 2
morphemes (what + is, where + is). What is particularly interesting from our
point of view is that this development appears to occur at dierent times in
dierent constructions.
We can see this when we compare contraction rates in dierent contexts at
dierent points in time. Ideally, one would want to follow changes in contrac-
tion rates in each question construction (e.g. what object questions with the
auxiliary is, what subject questions with the copula, etc.) but this would
require much larger corpora. With the data available, we can only follow the
combination of the most frequent WH words with the most frequent con-
tracted form, s. However, even when we collapse across dierent construction
types in this way, some patterns are clearly discernible. Figure 9.3 shows the
percentage of contracted forms after what and where in all three children and
after who in Naomi and Sarah (there is not enough data on who in Shem). The
ages on the X axis represent the midpoints of each period; and the nal point
represents the proportion of contracted forms in the input (i.e. the hypothet-
ical end result of development).
As we can see from the gure, adults used contracted forms equally fre-
quently after all question words, which suggests that any dierences in the chil-
drens levels of provision of such forms are not attributable simply to
phonological factors. There was some variation in the parents overall levels
of contraction, which presumably reect dialect dierences; however, within
the input to individual children, the levels of provision with dierent WH
words were very similar. Also, all three children, as well as their parents, con-
tracted after the pronoun that at the same levels (93 per cent or more of the
opportunities) throughout the entire period of observation. (This is not shown
in the gure to avoid cluttering it up.) On the other hand, the children used
contracted forms with dierent frequencies after dierent WH words, and,
what is more, the use of such forms developed along dierent trajectories in
dierent constructions.
The childrens questions with what follow a very similar developmental
pattern. All three initially preferred the contracted form in this context, and
supplied it about as frequently as their parents. This was followed by a period
of relatively frequent use of the full form, and later still, one presumes, by a
return to adult-like levels of provision, although we do not see this in the data
(except in Shem, where there is a slight rise at the last data point). Note,
however, that while the overall pattern of development is similar in all three
children, there are considerable dierences in the timing: the trough in the fre-
quency of contracted forms occurred about 2;9 in Shem and about 4;9 (or
perhaps even later, since this is the last data point) in Sarah.
Contraction after where, on the other hand, follows a dierent course of
development in each child. Sarah supplied contracted forms at about the
same rate throughout the entire period of observation. In Shem, we see a slow
but steady decrease in the frequency of such forms which continued even
when the frequency of contracted forms after what began to rise. Finally, the
Figure 9.3 Contraction of auxiliary/copula is after what, where and who
Naomi data shows a small decrease about 2;6, followed by a rise; note that
the frequency of contracted forms after what fell steeply during the same
period.
Contraction after who also follows a strikingly dierent pattern in the two
children for whom enough data is available. There are few questions with the
contracted form in Sarahs earlier transcripts; but after 3;6, her use of this
variant rose sharply to adult-like levels. Naomi, in contrast, consistently used

the contracted form during the entire period of observation.
Thus, we nd dierences in the pattern of use or non-use of contraction in
dierent constructions in the same children, as well as considerable dierences
between children. This supports the hypothesis that knowledge about auxil-
iaries develops independently in dierent constructions.
3.4.3 Errors
Further evidence for piecemeal learning is to be found in the childrens errors.
Errors specically, errors of commission are of course prime evidence for
productive language use, since they involve the production of structures that,
by denition, do not occur in the input. However, this productivity is for the
most part very restricted. Most errors of commission involve agreement or
auxiliary placement. The former kind arise most often when a plural NP is
inserted into a frame containing a singular verb (e.g. wheres X? + deir feets =
wheres deir feets?). Auxiliary placement errors occur when a WH word is com-
bined with a clause containing an auxiliary (e.g. where X? + hes taking it? =
where hes taking it?).
Most other errors are very restricted in scope. The Shem corpus, for
example, contains 20 double-auxiliary errors, in which the auxiliary verb is
placed both before and after the verb. Twelve of these are one-o innovations
involving a variety of dierent combinations of auxiliaries, e.g. did . . . is (what
did hes showing?), does . . . be (where does these pictures be taken?), is . . . are
(why is dese clothes are blowing by?). The remaining eight, however, involve
exactly the same combination of words: what i(s) hes doing? All of these
questions have he as subject and doing as the main verb, the rst auxiliary is
rendered as [] or [z] and the second is always contracted.
These questions are productive in the sense that they are extremely unlikely
to be imitations of an adult utterance; but the fact that the same mistake recurs
again and again in the same combination of words, but not with other sub-
jects and other verbs, suggests that they are formulaic in other words, it
seems that Shem memorised his own creative utterance and reproduced it on
subsequent occasions.13
A similar example can be found in the Naomi corpus. Between the ages of
2;0 and 2;3, Naomi produced 15 WH questions in which the contracted form
of the auxiliary is was followed by a bare innitive. All of these had the form
Whats Nomi do? With all subjects other than Naomi, she consistently used the
correct form of the verb: Whats Mummy doing?, Whats sh doing?, Whats
recorder doing? and so on.
Even more striking is Naomis use of forms of self-reference. Like many
children, she sometimes referred to herself using the rst person pronoun and
sometimes used her name (which she pronounced Nomi). These two forms,
however, occurred in dierent contexts. Naomi consistently used I in declara-
tives (99 per cent of the time) and intonation questions (100 per cent) as well
as questions with the modal auxiliaries can and could (also 100 per cent). On
the other hand, all but one of her 43 WH questions about herself contain her
name. They are all instantiations of three lexically specic formulas: Whats
Naomi do?, Whats Nomi doing? and Wheres Nomi? It is dicult to see how
such very specic restrictions in distribution could be explained without
appealing to preconstructed phrases.
3.5 Questions with long-distance dependencies

The schematic templates hypothesised above can account for Y/N questions
of arbitrary complexity and for WH questions about any element of a simple
clause. What is potentially problematic for such an account, however, is ques-
tions involving long-distance dependencies (LDDs), i.e. those in which the
questioned element is in a subordinate clause, as in the following examples:
(15) a. What did you say Chris needed?

b. What did you say they thought Chris needed?
c. What did you say they thought Maria believed Chris needed?
Here what is construed as belonging with the last verb in the sentence, namely
needed. Because there can be any number of clauses between the question
word and the place where it logically belongs, it is dicult to account for such
sentences using schematic templates: indeed, sentences with LDDs were the
main motivation for postulating WH movement in the rst place.
Not surprisingly, questions involving LDDs are acquired late. As far as we
can tell from the transcripts, only one of the three children whose linguistic
development was discussed earlier in this chapter, Sarah, produced any such
structures only three in total, all recorded at age 4;4 or later. Given that such
questions are relatively infrequent, their absence in the transcripts of the other
two children could be an accident of sampling. However, Table 9.9, which con-
tains the relevant gures for ve other developmental corpora, suggests that
this is not the case. As we can see from the table, only two of these additional
children produced the relevant structure, and in both cases it appeared fairly
late: at 3;8 in the case in the case of Abe, who was a rather precocious child,14
and at 4;3 in Adams case. The late emergence of questions involving LDDs is
conrmed by experimental research. In a series of production studies, Crain
and Thornton (Crain and Thornton 1991; Thornton and Crain 1994) were
able to elicit this structure from most 4-year-olds, but only a few younger chil-
dren succeeded on the task.
It is also worth noting that the questions with LDDs were quite formulaic.
Of the 49 tokens produced by the children (44 by Abe, two by Adam and three
by Sarah), all but two had the form WH do you think S? There was some
variety within this schema, in that the children used dierent WH words to ask
about dierent parts of the subordinate clause.
Table 9.9 Child questions with long-distance dependencies (LDDs) in five

other CHILDES corpora
Corpus Corpus size Age LDD Age of
(child words) range questions emergence
Manchester (Theakston et al. 2000) 538,000 1;83;0 0

Hall (Hall et al. 1984) 295,000 4;65;0 0
Gleason (Belliger and Gleason 1982) 60,000 2;15;2 0
Adam (Brown 1973) 164,000 2;34;10 2 4;3
Abe (Kuczaj 1976) 163,000 2;45;0 44 3;8
Note: The first three corpora contain samples from a number of children; the last
two from individual children. The figures include only questions containing a WH
word followed by an auxiliary and subject and a finite subordinate clause.
(16) a. subject: What do you thinks under here? (Abe, 3;10)

b. locative complement: Where do you think theyre going?
(Abe, 3;11)
c. adjunct: How do you think she made that train? (Abe, 4;11)
d. subject predicative: What do you think it was? (Abe, 4;0)
All but two of the questions, however, had think as the main verb, you as the
main verb subject, and do as the auxiliary. The same is true of the questions
elicited in the Crain and Thornton studies. Thus, by about age 4, children have
some knowledge about questions with LDDs; but it is not clear how general
such knowledge is.
Interestingly, adult LDD questions are also highly stereotypical. There are
325 tokens of this construction in the adult input in the Manchester corpus.
None of these contains more than one nite subordinate clause that is to say,
there are no sentences such as examples (14b) and (14c) above and only one
has an overt complementizer. In 96 per cent of the questions, the main verb is
either think or say; 91 per cent have you as subject (and 6 per cent have other
pronominal subjects); and 99 per cent have some form of do in the auxiliary
position. Moreover, specic main verbs appear to be associated with specic
auxiliary forms: think with do (276 out of 280 tokens), say and tell with did (31
out of 33 and six out of six tokens respectively), and pretend with shall (three
out of three tokens). The remaining four verbs (see, hope, reckon, suppose)
occurred only once in LDD questions, and hence could not occur with dier-
ent auxiliaries. In fact, even the verbs in the subordinate clauses are fairly
stereotypical, so that the following eight templates account for 50 per cent of
all LDD questions in the entire corpus (CAPITALS indicate that a particular
verb occurred in several dierent forms: for example, BE stands for is, was, are,
could be etc.):
(17) a. What do you think NP BE?

b. What do you think NP DO?
c. What do you think NP EAT?
d. What do you think NP LIKE?
e. What do you think NP NEED?
f. Where do you think NP BE?
g. Where do you think NP GO?
h. What did NP say NP was?
Arguably, child-directed speech is not the most representative example of

adult linguistic abilities, so it is important to obtain comparable statistics
about other genres. The relevant evidence is oered by Verhagen (forthcom-
ing), who studied LDD questions in the Brown corpus (which includes a wide
variety of written texts) as well as two Dutch corpora. The Brown corpus con-
tains 11 instances of questions involving LDDs. In 10 of these, the matrix verb
is think, and in one, say; in nine, the subject is do, while the remaining two take
third person pronouns as subjects; and 10 contain some form of do. The
Dutch results are very similar: the overwhelming majority of LDD questions
have second person subjects and denken think as main verb. Verhagen also
notes that when an LDD question deviates from the prototype, it does so in
only one respect: that is to say, it either takes a dierent matrix verb or a dier-
ent subject, but (in his data) never both.
Similar observations can be made about the Manchester corpus. Of the
questions with LDDs, 89 per cent instantiate the WH DO you think/say S
prototype; 9 per cent deviate from prototype in one respect (they contain a
dierent matrix verb or a dierent auxiliary or a dierent subject); and only 2
per cent deviate from the prototype in more than one way. This clustering
around a prototypical pattern suggests that speakers may produce such utter-
ances by inserting novel material into a ready-made template and modifying
it when necessary (e.g. by substituting a full NP for you) rather than assem-
bling them from scratch. This possibility will be discussed more fully in the
next chapter; for now, we will examine some independent evidence that adult
speakers store partially lexically specic templates of long-distance WH ques-
tions.
The evidence comes from a grammaticality judgement study involving
prototypical and unprototypical LDD questions. The former had the form
WH do you think S-WH? or WH did you say S-WH?, where S-WH stands for
a subordinate clause with a gap. Unprototypical questions contained a dier-
ent auxiliary (will, should, would, might or have), a lexical subject (e.g. the cus-
tomers, your sister), and a dierent main clause verb (e.g. claim, reveal, believe,
remember), followed by the same subordinate clause. Interspersed with these
20 test sentences were 20 declarative control sentences containing the same
lexical material and 10 ungrammatical llers (ve interrogatives and ve
declaratives; see Table 9.10 for examples). The participants (10 adult native
Table 9.10 Examples of sentences used in the grammaticality judgement

study
Condition Example
Prototypical interrogative Where do you think they sent the documents?

Unprototypical interrogative Where will the customers remember they sent the
documents?
Prototypical declarative You think they sent the documents to the Head
Oce.
Unprototypical declarative The customers will remember they sent the
documents to the Head Oce.
Ungrammatical *Who do you think that left?
*He left in the oce and had to go back.
Table 9.11 Acceptability ratings

Condition Mean SD
Prototypical interrogative 4.5 0.4

Unprototypical interrogative 3.3 0.6
Prototypical declarative 4.0 0.5
Unprototypical declarative 3.7 0.5
Ungrammatical 1.9 0.6
speakers of English) were asked to rate the acceptability of the 50 sentences

on a scale from 1 (completely unacceptable) to 5 (fully acceptable).
If people rely on fully general rules, there should be no dierences between
prototypical and unprototypical questions, provided that the corresponding
declaratives are equally acceptable. If, however, they use lexically specic tem-
plates, they should rate the prototypical questions as more acceptable than the
unprototypical ones. The mean acceptability ratings and standard deviations
for each type of sentence are given in Table 9.11. As we can see, prototypical
questions were indeed given higher ratings than unprototypical questions. The
eect is quite robust and highly signicant (t(9) = 7.54, p < 0.001). The dier-
ence in the acceptability of the corresponding declaratives, in contrast, was
not signicant (t(9) = 1.57, p = 0.15), showing that the eect cannot be attrib-
uted to the combination of the lexical properties of the auxiliary, subject and
verb in the unprototypical sentences. Thus, the results support the hypothesis
that people store lexically specic templates for long-distance questions. It
should be noted, however, that the ratings for unprototypical questions were
considerably higher than those for the ungrammatical controls (t(9) = 5.93, p
< 0.001), which suggests that speakers also have some more general knowledge
about such structures.
3.6 Conclusion
On the standard generative account, the acquisition of questions involves
learning to apply two rules, WH fronting and subjectauxiliary inversion. Both
of these rules are very general, but subject to certain constraints. Such con-
straints are thought to be unlearnable, and hence the fact that humans appar-
ently obey them is sometimes regarded as providing the clinching evidence for
the existence of innate Universal Grammar (Cf. chapter 6, section 4.1).
The evidence reviewed in this chapter, however, suggests a very dierent view
of development. We saw that early usage is highly stereotypical and that devel-
opment proceeds from invariant formulas through increasingly general formu-
laic frames to abstract templates. Furthermore, development is piecemeal, in
the sense that children learn the correct word order independently for each con-
struction,15 and the same changes occur at dierent times in dierent con-
structions. We also found no evidence of abrupt changes in behaviour which
might indicate a shift to a dierent productive mechanism: instead, we see slow,
steady progress. This suggests that the endpoint of development that is to say,
adult grammar might be rather similar to the kinds of representations that
children use; in other words, that adult knowledge comprises not constraints
on movement and suchlike, but rather fairly general templates like Can NP
VP?, Who VP? and What BE NP TRANS.V-ing? These would also allow
speakers to produce and understand a wide variety of dierent question types,
thus accounting for the observed exibility of adult behaviour. However, they
do not involve movement, and hence do not require innate constraints or an
abstract level of syntactic representation distinct from surface structure.
Some linguists have argued that speakers ability to produce and understand
questions involving LDDs constitutes evidence against schematic templates,
since such dependencies are unbounded: that is to say, there is no limit (in
principle) on the number of clauses intervening between the WH word and the
place where it logically belongs. Clearly, an adequate account of LDD ques-
tions must explain how such structures are produced and understood.
However, our examination of adult use of such structures revealed that they
are in fact highly formulaic, so the data actually supports the hypothesis that
even adults rely on lexically specic templates.
It should be stressed that the use of such frames requires grammatical
knowledge. Speakers must know that certain slots can only be lled with items
which have certain grammatical properties (for example, the VP in who VP?
must be tensed, while the VP in can NP VP? cannot); and there are often rela-
tionships between items in dierent grammatical slots (for example, in What
BE NP TRANS.V-ing? the auxiliary must agree with the subject in person and
number). The challenge for linguistics, then, is to develop a theory of grammar
which will capture such knowledge in psychologically realistic terms. The next
chapter introduces one such theory.
Notes
1. The processing complexity and formulaic explanations of processing

reversals, and the other proposals mentioned earlier, are not mutually
exclusive. The use of prefabricated patterns does not preclude tinkering
with them to adapt them to ones present communicative goals; and such
tinkering is less likely to occur in cognitively demanding situations.
2. To discover the shifting nature of deictic pronouns, children need to
observe speech events involving third parties, since in speech addressed to
them, you always refers to the child and I to the caretaker. Since blind chil-
dren often do not know what these third parties are doing, discovering the
deictic properties of the pronouns is much more dicult.
3. A similar case of a childs developing a phonological form which does not
occur in adult speech but which can be viewed as a phonological common
denominator of several adult forms expressing a similar meaning is pro-
vided by Mills (1985). The rst denite article used by some German-
speaking children is de, which is the nearest you can get to the common
denominator of the adult forms der, die, das, dem, des and den.
4. A researcher working in Menns lab compared audio tapes of Berko
Gleasons subject Patricia with transcriptions available on the CHILDES
database. Patricia rendered the object pronoun it as /t/, // or zero, forcing
the transcribers either to underrepresent or overspecify Patricias knowl-
edge. They chose the latter (Peters and Menn 1993: 753, footnote). Such
idealisations give a distorted picture of grammatical development,
making it appear less continuous than it really is.
5. Sentences with tags of matching polarity (He is coming, is he?) are an
instance of a dierent construction with a dierent pragmatic function
and a distinct intonation.
6. Unless we have a full record of everything the child said, we have no way
of knowing the actual frequency of any expression in his or her speech.
It is possible that the three recorded occurrences were the only ones the
child ever produced; and it is also possible that an utterance with a fre-
quency of three in the corpus had actually been produced hundreds of
times. However, assuming the sample to be representative, the mean
actual frequency of expressions with a frequency of three in the sample
is 30 if we have a 10 per cent sample, and 43 if we have a 7 per cent
sample.
7. Errors of commission are errors involving incorrect use (e.g. wrong form
or wrong word order), as opposed to errors of omission, which can be cor-
rected simply by adding the omitted elements.
8. It is possible that Shems development showed a U-shaped pattern similar
to Naomis, but we dont see this in the data because the recordings began
fairly late. If this is the case, then the trough would be much deeper in
Shems case than in Naomis (notice that he inverted less than 20 per cent
of the time at the beginning of the recording period, while Naomis lowest
rates were about 40 per cent).
9. Naomis mothers questions were used for this comparison. The data ana-
lysed here consists of a corpus of the mothers questions matched in size
to the childs corpus.
10. There are only four Y/N questions with do in the rst half of the Sarah
corpus, so this part of the data was not included in Table 9.6.
11. The dierences are not statistically signicant in the mother: 2 = 1.2, df
= 2, p = 0.550. However, they are signicant in all three children: for
Naomi 2 = 12.6, df = 2, p = 0.002; for Shem, 2 = 31.7, df = 3, p < 0.001;
for Sarah2, 2 = 8.3, df = 3, p = 0.039.
12. For the mother, 2 = 4.2, df = 3, p = 0.236; for Naomi, 2 = 20.6, df = 3,
p < 0.001; for Shem, 2 = 27.3, df = 3, p < 0.001; for Sarah1, 2 = 13.5, df
= 3, p = 0.004; for Sarah2, 2 = 28.0, df = 3, p < 0.001.
13. Note that hes was not an unanalysed chunk. Shem did not use it in con-
texts which did not require an auxiliary or copula or in combination with
an auxiliary in declaratives.
14. Abes rst recorded use of a question with an LDD occurred at age 2;10,
but this was an imitation of the immediately preceding adult utterance.
15. It remains to be seen whether this is also true of later stages of language
acquisition. Development does speed up noticeably at about age 3, sug-
gesting that there may be transfer of knowledge between constructions
during this period.
10 The cognitive enterprise
1. Towards a psychologically realistic grammar
In the rst part of this book, we looked at some general properties of human
language that must be accommodated in any model of language that purports
to be psychologically realistic. Language processing is exible (we are able to
make use of whatever information is available and integrate various sources of
information as required) and very fast. It is also remarkably robust, in that it
develops even in very inauspicious circumstances and is dicult to destroy
completely, so that some linguistic abilities are preserved even after damage to
the parts of the brain which normally play a central role in language process-
ing. Last but not least, it relies to a large extent on relatively low-tech, general-
purpose mental mechanisms shared with other species, including the ability to
perceive, categorise and store information, form cross-modal associations,
and plan hierarchically organised action sequences.
In Part II so far, we have examined specic aspects of linguistic organisa-
tion in more detail, considering developmental evidence as well as research on
adult knowledge. This has enabled us to glean some hints about how these
properties are implemented in the language machine. One of the reasons why
language processing is so fast is because it relies to a considerable extent on
prefabricated units which need to be modied only minimally in order to be
used in communication. The exibility and robustness of the system are due,
in a large measure, to its redundancy: the fact that the same information is rep-
resented in dierent parts of the system, and hence there are typically dier-
ent ways of assembling the same utterance or arriving at its meaning. These
are tried out in parallel, which also makes for speed. We have also seen that
memory plays a very important role in language acquisition and production.
There is an intimate relationship between lexical and grammatical develop-
ment, between rote learning and rule extraction; and grammatical phenom-
ena, like words, are subject to frequency eects.
Much of the research discussed in Part II so far has focused on language-
specic aspects of linguistic knowledge. We have seen that such knowledge
can be quite subtle. It follows that we need a well-articulated descriptive
framework to capture it, and a theory of acquisition which explains how humans
are able to learn the idiosyncratic features of their language as well as the more
general aspects which can be captured by means of universal principles.
This chapter is devoted to approaches to language which meet these require-
ments. While the requirements place strong constraints on linguistic theory,
they still allow for a number of alternatives. Rather than surveying all the pos-
sible options, I will focus on one particular theory, Cognitive Grammar
(Langacker 1987, 1991a, 1991b, 2000), which is the most fully elaborated of
these, although the discussion will also incorporate some key ideas from
related frameworks, notably Construction Grammar (Fillmore and Kay 1993;
Fillmore et al. 2003; Goldberg 1995) and Radical Construction Grammar
(Croft 2001). I will begin with an informal overview of this framework, and
then show how the kinds of representations that it postulates might develop
and be used in production. The chapter concludes with some general observa-
tions about the challenges facing cognitive linguistics.
2. A crash course in Cognitive Grammar
2.1 Linguistic expressions are symbolic units

The central idea of Cognitive Grammar (CG) is that language including
grammar is essentially symbolic in nature, and thus linguistic knowledge can
be captured by means of symbolic units, or pairings of a phonological form
and a meaning. Figure 10.1a illustrates the usual conventions for representing
symbolic units. The schematic diagram in the top panel stands for the meaning
of the lexeme cat (or, in Langackers terminology, its semantic pole) and the
phonemic transcription in the bottom panel for its pronunciation, or phono-
logical pole; the vertical line connecting the two panels corresponds to the
symbolic relationship between the two. The caption underneath the gure
Figure 10.1 The symbolic units []/[kt] (a) and [[]-]/[[kt]<[s]] (b).
Note: Boxes indicate semantic and phonological units; vertical lines correspond to
symbolic relationships. The < symbol indicates linear order.
The cognitive enterprise 205
shows an alternative way of representing the same unit: by convention,

are used to abbreviate the meaning of an expression, and phonemic
transcription its phonological form. The slash indicates that the latter symbol-
ises the former.
Symbolic units can be simple, as in the above example, or complex. The
latter include morphologically complex words (e.g. cats, relationship), phrases
(after school, in the long run, deep down) and sentences (Alls well that ends well,
Would you like a cup of coee?): all of these expressions, like simple lexemes,
are pairings of a phonological form and a meaning. Such complex units
consist of two or more components, each of which is a symbolic unit in its own
right. For example, the symbolic unit cats is made up of two simpler units, the
noun cat and the plural morpheme -s (Figure 10.1b). Symbolic units can also
be partly or wholly schematic. As we shall see shortly, schematic symbolic
units are used in the CG framework to capture grammatical knowledge.
2.2 Imagery
The raison dtre of linguistic forms is to express meaning. Cognitivists take a
rather broad view of meaning as comprising not just content but also
imagery. This refers to our ability to construe the same content in alternate
ways by focusing on dierent aspects of a scene, viewing it from dierent per-
spectives, or conceptualising it at dierent levels of abstraction.
One of the most ubiquitous aspects of imagery is the imposition of a prole
on a base. The base is the knowledge structure within which an expression is
understood and dened; it is similar to Fillmores notion of frame (Fillmore
1975, 1978) and Lakos idealised cognitive model (Lako 1987). The prole
is a substructure within the base that an expression actually designates. For
example, the expression diameter designates, or proles, a line, but it also
evokes a larger cognitive structure a circle which is its base: it is impossible
to entertain the notion of without at the same time entertaining the
notion of . Two linguistic expressions can have the same base but dier-
ent proles and hence contrast in meaning. For example, the terms aunt and
niece designate dierent nodes in the kinship network which is their base; win,
lose, and beat prole dierent relationships in a knowledge structure we could
describe as a ; and winner, loser and runner-up prole indi-
viduals participating in these relationships. Conversely, two expressions can
have the same proles, but dier in meaning because the proles are imposed
on dierent bases. An example would be the adjective dry and the past parti-
ciple dried: both prole the same property (lack of moisture), but only the
latter has as its base a series of events in the course of which an object under-
goes a change of state (from wet to not wet).
Another dimension of imagery, crucial to the conceptualisation of rela-
tions, is trajector/landmark organisation. When we perceive relations, we tend
to impose a certain perspective on the participating entities. Specically, we
tend to view one of the participants as the central element of a scene a gure
that stands out against a background of other entities. In CG, the most salient
entity is called the trajector and the other salient participants are called land-
marks. For instance, in (1), the Toyota is the trajector and the Mercedes the
landmark; in (2), the roles are reversed. In this example, the relationship
between the trajector and the landmark is spatial, but the terms are used to
refer to the corresponding roles in other kinds of relations as well. Complex
linguistic expressions usually have several levels of trajector/landmark organ-
isation. In (3), Chris is the trajector of put and the drink its landmark; but the
latter is also the trajector of on, while the table is its landmark.
(1) The Toyota is in front of the Mercedes.

(2) The Mercedes is behind the Toyota.
(3) Chris put the drink on the table.
In most cases, some element of the scene is inherently more salient than the
others and hence the most natural choice for the trajector. For example,
moving objects and agents are more likely to be construed as trajectors than
stationary objects or patients. This prototypical alignment is reected in (3),
where the agent is the clausal trajector and the displaced object the trajector
of the preposition. However, even when the relationship is inherently asym-
metrical speakers enjoy a certain degree of freedom and may choose the less
likely candidate as the gure. The humorous eect of (4), for example, is due
largely to the fact that the displaced object (the Scotch and soda) is construed
as landmark of outside, while the most natural choice for the landmark (the
container into which the Scotch is poured) becomes the trajector.
(4) He was seated at the Savoy bar when he told me this, rather feverishly
putting himself outside a Scotch and soda. (P.G. Wodehouse, The
World of Mr Mulliner)
Or consider a scene in which two cars, a Toyota and a Mercedes, crash into
one another. Although all of the sentences in (5) could be used to describe this
event, they are clearly not synonymous. Choosing one of the cars as the tra-
jector has the eect of highlighting its causal role in the crash and downplay-
ing the role played by the other car; construing both cars as trajectors results
in a more neutral description.
(5) a. The Toyota collided with the Mercedes.

b. The Mercedes collided with the Toyota.
c. The Toyota and the Mercedes collided.
Similar observations can be made about the pairs of sentences given in (6)
and (7). In both cases, the two sentences describe dierent construals of the
same event and hence dier in meaning, although the dierences are quite
subtle. Since the content words are the same in both members of each pair, it
is clear that the dierence in meaning is part of the meaning of the construc-
tion and the grammatical functors that form an integral part of it.
(6) a. Lil decorated our wedding cake.

b. Our wedding cake was decorated by Lil.
(7) a. Andy sent Neil the new personal organiser.

b. Andy sent the new personal organiser to Neil.
Yet another aspect of imagery is our ability to conceive the same content at
dierent levels of schematicity. For example, we can think of the same person
as , , or , or of the same activity as
, or . We are also able to make
categorising judgements such as (the concept is sche-
matic for ) and .
The term schema refers to the superordinate term in such relationships; the
subordinate term is called the schemas instantiation. These terms are more
general than the traditional labels superordinate and hyponym in that they
apply to all kinds of linguistic units, including phonological units (for
example, CVC is schematic for [kt], [nid], [f] and so on) and grammatical
units (see below).
2.3 Things, processes and atemporal relations

The distinction between content and imagery is extremely useful in providing
a semantically based account of grammatical distinctions which in other the-
ories are dealt with by means of arbitrary syntactic features. Consider the
expressions investigate and investigation. As far as actual semantic content is
concerned, the two words are identical; but they contrast semantically because
they impose dierent images on the same content. Investigate proles a
process, or a series of states occurring at dierent points in time; investigation,
on the other hand, involves conceptual reication, or construing an event as a
type of thing. Thing in CG is a technical term dened as a region in some
domain (Langacker 1987: 189), where a region is a set of interconnected enti-
ties (Langacker 1987:198). The term covers not just physical objects, but
various abstract entities as well, including stretches of time, colours and qual-
ities. Thus, week designates a region in the temporal domain; the noun blue
designates a region in colour space; beauty designates a region in quality
space; and so on. Even events can be construed as abstract things in this tech-
nical sense. The component states of an event include the same participants
and occur at adjacent points in time. Because of these interconnections, the
component states constitute an implicit region. A nominalisation such as
investigation proles this region (and backgrounds the temporal dimension).

Thus, the noun and the verb impose contrasting images on the same content:
they have the same base, but dierent proles (Langacker 1991b: 97).
Lexical categories such as nouns and verbs are also considered symbolic
units in CG, albeit highly schematic ones. A noun is an expression which pro-
les a thing, with maximally underspecied phonological content (that is to
say, it must have some content, but it can be any speech sound whatsoever). A
verb is dened as a phonologically unspecied symbolic unit which proles a
process, or a temporal relation; and prepositions, adjectives and adverbs
prole atemporal relations (Langacker 1991b: 78).
2.4 Constructional schemas

As indicated earlier, symbolic units can be simple or complex and concrete or
schematic. A symbolic unit which is both complex and schematic is called a
constructional schema. Constructional schemas are the CG equivalent of
rules: they capture speakers knowledge about the internal structure of
complex expressions, and can be used to assemble novel expressions. However,
unlike traditional rules, they are represented in the same format as their
instantiations (that is to say, actual expressions of the language).
This is best explained by means of an example. Figure 10.2 shows simplied
representations of the expression The girl wants a cat and two constructional
schemas, WANTER.NP want WANTED.NP and NP V NP. The expression
The girl wants a cat describes a process or temporal relationship involving two
participants, a girl and a cat, represented in the diagram by simplied draw-
ings. The relationship itself is represented by the horizontal line labelled
which connects the two participants. Although this relationship is stative (i.e.
it doesnt change in time), it is nevertheless a temporal relation, in that it
involves a series of congurations scanned sequentially (cf. Langacker 1987,
1991a, 1991b) observed at dierent points in time, as it were. Thus, the tem-
poral dimension is part of its prole; this is represented by the arrow in the
diagram. As in the previous gure, the lower panels represent phonological
form, and the vertical lines the symbolic relationships between the signiant
and the signi. Thus, the phonological units [
l], [
wnts] and [
kt]
symbolise the semantic units , and , respectively; and the
symbol < indicates the linear order of the three elements. (To simplify the dis-
cussion, I am ignoring the contribution of the determiners and the third
person inection on the verb.1)
Figure 10.2b shows a low-level schema representing the same relationship
at a more abstract level.2 The schema captures the commonality in various sen-
tences containing the relevant sense of want (The girl wants a cat, Her broth-
ers want three dogs and a rabbit, Their mother wants some peace and so on),
and hence the speakers knowledge about the combinatorial possibilities of
this verb. It can also be regarded as (part of) the lexical entry for want. The
Figure 10.2 A simple clause (a), a low-level schema (b), and an abstract schema (c)
semantic pole of the schema, like that of the concrete expression, contains a
subpart corresponding to the trajector (the wanter) and the landmark (the
desired object), since they are part of the meaning of want (it is impossible to
think of wanting without simultaneously thinking of an animate being expe-
riencing this state and the object of their desire). Both of these participants
are things in the technical CG sense. The phonological pole spells out the seg-
mental content of the predicate and the linear order of the expressions corre-
sponding to the wanter and the desired object in relation to the predicate; the
actual segmental content of the expressions symbolising the trajector and
landmark is unspecied (this is represented in the diagram by three dots, which
act as place markers, or slots into which contentful expressions can be
inserted).
Finally, Figure 10.2c represents the abstract transitive construction, which
is highly schematic at both levels: the semantic representation indicates that
the construction proles a temporal relationship involving two things, and
the phonological representation spells out the linear order of the three com-
ponents. The actual segmental content of these component units is left
unspecied.
The three diagrams in Figure 10.2 represent the same content at varying
levels of abstraction: WANTER.NP want WANTED.NP is schematic for The
girl wants a cat, and NP V NP is schematic for both of these units. The crucial
point is that all three representations are symbolic units and have the same
overall structure. In this respect, the CG approach is very dierent from most
mainstream approaches to language, in which syntactic knowledge is repre-
sented in a dierent format from lexical knowledge, using constructs such as
phrase structure rules, movement rules, and various abstract principles and
constraints. In the CG framework, schemas and their instantiations are repre-
sented in the same format and dier only in the degree of specicity: the
schema is specied in less detail than its instantiations. It follows that schema
extraction involves reinforcement of shared properties and loss of detail,
rather than translation into a dierent representational format (see below).
2.5 Language as a structured inventory of conventional linguistic units

In the CG framework, linguistic knowledge is seen as a complex, richly inter-
connected network (or structured inventory) of linguistic units. There are
three types of units: phonological, semantic and symbolic, and three types of
relationships between them: symbolisation, composition and schematicity.
Symbolisation, as explained earlier, is the relationship between a chunk of
semantic structure and the phonological form which can be used to evoke it.
Composition relationships hold between component units and the composite
unit which they form, and are found at all three levels of linguistic organisa-
tion. For example, the noun cat and the plural morpheme are components of
the complex symbolic unit cats, which in turn is a component of higher-order
units such as its raining cats and dogs. Likewise, the phonological form [kts]
and the semantic unit [[]-] consist of simpler elements and can form
parts of larger assemblies. Schematicity is the relation between the schema (the
superordinate term in a taxonomic hierarchy) and its instantiation (the subor-
dinate term). As in the case of composition, schematic relationship can hold
between semantic units (e.g. is schematic for , which in turn is sche-
matic for ), phonological units (CVC is schematic for the phonologi-
cal form [kt], which is schematic for actual phonetic realisations of the word),
and symbolic units (the plural construction [[]-]/[S] is sche-
matic for the expression cats, and the abstract symbolic unit []/[] is
schematic for both cat and cats).
All the examples given above involve full schematicity: the instantiations are
fully compatible with the properties of the schema, but are characterised in
more detail. This is not always the case: a linguistic expression (or phonolog-
ical form or conceptualisation) may not fully match the properties of the
schema, but still be included in the category dened by the schema. In such
cases, we speak of partial schematicity or extension. This is typically the rela-
tion between the dierent senses of a polysemous word. Consider, for
example, one of the basic senses of the verb y (move through the air by ap-
ping the wings) and an extended sense meaning utter (as in Her hair was
ying in the wind). Both senses designate situations involving a particular kind
of movement in the air, but in the extended sense, the moving object is
attached to something and is not displaced.
We saw earlier that polysemy is rampant in language (see Chapter 2, section
1; Chapter 7, section 1.2; Chapter 8, section 9), so the semantic representations
of most words are not single senses but families of related senses. This is illus-
trated in Figure 10.3, which shows some intransitive uses of the verb y and
the relationships between them. Note that each sense is related to at least one
other sense, and that there are several schemas (represented by the three boxes
in the upper part of the diagram) which capture some of the local similarities:
move through the air, move quickly, resist the force of gravity. However,
there is no single superschema which would be compatible with all the uses of
y.
Most constructional schemas, too, are complex categories in which the
various meanings are connected by relations of full or partial schematicity (cf.
Goldberg 1995; Lako 1987). For example, the caused motion construction
has at least ve interrelated senses (Goldberg 1995):
(8) a. X Y / Z: Barbara dragged Michael out of

the kitchen.
b. X Y /
Z: Barbara ordered Michael out of the kitchen.
c. X Y / Z: Barbara let Michael out of the
kitchen.
26/3/04
4:35 pm
Page 212
Figure 10.3 Part of the semantic network for fly.

Note: Each box in the bottom part of the diagram represents an established sense of the verb. The three boxes at the top correspond to
local schemas. Horizontal lines indicate relationships of partial schematicity; vertical and diagonal lines indicate relationships of full
schematicity. Boxes with heavy lines represent strongly entrenched senses.
d. X Y / Z: Barbara locked
Michael out of the kitchen.
e. X Y / Z: Barbara helped Michael out of the
kitchen.
Senses (b)(e) share some similarities with the basic sense exemplied in (a)
which can be captured by means of local schemas. However, there are no fea-
tures common to all ve senses (apart from the fact that they describe some
kind of action; this, however, is a property shared with many other construc-
tions).
Finally, many words have variant pronunciations which share most but not
all features, and these form similar interconnected structures. Thus, most sym-
bolic units are not simply pairings of a phonological form with a meaning, but
rather pairings of a family of senses with one or more pronunciations sub-
networks within the larger network that constitutes the grammar of a lan-
guage.
2.6 A usage-based model

CG emphasises the importance of usage. Speakers knowledge of their lan-
guage is captured in phonological, semantic and symbolic units which are gen-
eralisations over actual usage events. It is also assumed that each usage event
leaves a trace in the processing system. Every time a particular unit is accessed,
its representation is strengthened, or entrenched, so units which are accessed
frequently become easier to activate; conversely, units which are not accessed
gradually decay, and become more dicult to activate. It follows that a
persons mental grammar is a dynamic system constantly shaped by experi-
ence with the language. Early in development, such changes are relatively
rapid and large; later modications involve smaller adjustments ne-tuning
of the system.
The CG view of language is also maximalist in the sense that linguistic
knowledge is represented at various levels of abstraction: concrete units exist
alongside schemas of varying degrees of abstraction. Moreover, CG empha-
sises the importance of low-level schemas, since these are necessary to capture
actual patterns of usage:
Lower-level schemas, expressing regularities of only limited scope, may on

balance be more essential to language structure than high-level schemas
expressing the broadest generalizations. A higher-level schema implicitly
denes a large space of potential instantiations. Often, however, its
actual instantiations cluster in certain regions of that space, leaving other
regions sparsely inhabited or uninhabited altogether. An adequate
description of linguistic convention must therefore provide the details of
how the space has actually been colonized. (Langacker 2000: 29)
It follows that in ordinary language use, speakers rely preferentially on con-

crete lexical units and local schemas that is to say, units which are relatively
close to actual usage events. Our mental inventories also contain more abstract
schemas capturing higher-level generalisations, but these, Langacker suggests,
have an organizing function rather than an active computational one (cf.
Langacker 1991b: 265).
2.7 Meeting the specications

It should be evident even from this brief introduction how CG meets the
requirements outlined earlier. The CG conception of linguistic knowledge as a
network of symbolic units of varying size and degree of schematicity can easily
accommodate prefabs. The network is highly redundant and richly intercon-
nected, so speakers can access the information represented in it in various ways.
For example, a complex lexical expression such as Alls well that ends well might
be accessed in a single step as a ready-made unit. Alternatively, the speaker
might rst activate some salient components (e.g. the lexical units end and well),
which in turn will activate the composite unit; and a speaker who has never
heard the expression will have to construct it from scratch, without using any
prefabs. These features help to explain the observed exibility and robustness of
language, and also allow for the possibility that individual speakers mental
grammars dier in many details. Furthermore, since the nodes in the network
dier in entrenchment, the cognitive approach can easily accommodate fre-
quency eects. Last but not least, CG relies whenever possible on cognitive
mechanisms which are not specically linguistic, that is to say, mechanisms
involved in perception (e.g. perceptual primitives, gureground organisation),
categorisation and memory. It does not rule out the possibility that humans are
endowed with language-specic biological adaptations: however, it is assumed
that these are likely to be fairly modest, and should only be postulated as a last
resort, when there is no other way of explaining a given phenomenon.
Another extremely attractive feature of the CG approach is the fact that
lexical and grammatical knowledge is captured in the same representational
format, with no sharp contrast between grammar and the lexicon. As pointed
out earlier, constructional schemas dier from phrasal lexemes only in degree
of specicity; and lexical entries for verbs and other relational expressions are,
in eect, mini-grammars. Thus, the close links between grammatical and
lexical knowledge observed in language acquisition and impairment fall out
as a natural consequence of the organisation of the network.
3. Language production in a CG framework
Linguistic communication involves the use of symbolic expressions, which are

pairings of a form and a meaning. Both participants begin with half of the
expression (either the form or the meaning), and their task is to (re)construct
the other half: thus, production basically involves nding a phonological form
which matches ones communicative intentions, and comprehension, nding a
conceptualisation which matches the sounds one is hearing. In the simplest
case, a speakers mental lexicon will contain a unit with the appropriate
meaning, so all the speaker needs to do is to access its phonological form and
pronounce it. If the expression is also available to the listeners as a unit, all
they have to do after they have recognised it is access its semantic representa-
tion. On most occasions, however, a ready-made unit will not be available. In
such cases, the speaker will have to nd units which correspond as closely as
possible to parts of his or her communicative intentions and combine them
into a novel expression; and the listener will have to recognise the component
units, access their semantic representations, and integrate them into a coher-
ent conceptualisation. For the sake of simplicity, in the following discussion
we will look at the processing of novel expressions from the perspective of the
speaker; but many of the observations about the operations involved in sym-
bolic integration also apply to comprehension.
New expressions can be assembled by juxtaposing or superimposing estab-
lished units. Juxtaposition involves simply putting the two components next to
each other (in either order). For example, juxtaposing come here and now yields
come here now or now come here; juxtaposing Kate and would you like another
drink? yields either would you like another drink, Kate? or Kate, would you like
another drink? The relationship between the two components is paratactic
rather than truly syntactic: the fact that they are placed next to each other
signals that their meanings are to be integrated, but the grammar does not spell
out how this is to be done, so it must be inferred by the listener. Furthermore,
the grammar doesnt impose any requirements on the component expressions
(other than the fact that they should be elements capable of functioning as
utterances in their own right, given the right context). For example, now can
combine with a clause or a verb phrase (as in the examples above), with a prep-
ositional phrase (in America now) or just a noun phrase (now the gravy).
Superimposition, on the other hand, is a more constrained process in which
one component (a ller) elaborates a schematically specied subpart of
another component (the frame). For example, the expressions want and my
desk can be superimposed to derive the composite expression want my desk (cf.
Figure 10.4). Superimposition happens simultaneously at the phonological
and the semantic poles of the two components. In Figure 10.4, this is shown
by the dotted lines linking corresponding elements: elaborates the
landmark of the semantic structure [[] [ ]], and the
phonological form [ma
desk] elaborates the second slot in [wnt]. (Again,
the diagram is simplied: for example, it does not represent the internal struc-
ture of my desk.) Note that the ller must match the properties specied in the
frame: in our example the frame requires a ller designating a thing (in the
technical CG sense).
Figure 10.4 Superimposition of a typical frame and filler
Figure 10.5 Superimposition with mutual elaboration
It is also possible for both units to be partially schematic and elaborate

dierent parts of each other: in such cases each unit acts as a ller as well as
a frame. This is illustrated in Figure 10.5, which depicts the superimposition
of the symbolic units want and them. Again, integration takes place simulta-
neously at the semantic and the phonological poles of the expressions. At the
semantic pole, 3 (third person plural) elaborates the schematically specied
landmark of [[] [ ]], and elaborates the sche-
matically specied process in the semantic representation of [[]
[3]]. At the phonological pole [wnt] elaborates the second slot in [
m] and at the same time [m] elaborates the second slot in [wnt].
In subsequent discussion, I will treat phonological and semantic superim-
position as a single process of symbolic integration, and will use italics to indi-
cate symbolic expressions, with traditional grammatical labels (NP, V etc.)
representing phonologically underspecied components in symbolic units, or
slots which can be elaborated by more contentful units. For example,
WANTER.NP want WANTED.NP and NP V them represent the symbolic
units [[]--[ ]]/[wnt] and [[]

[3]]/ [ m] respectively.
It was suggested earlier that speakers store units of various sizes and degrees
of specicity: general templates such as NP V NP, low-level schemas such as
WANTER.NP want WANTED.NP, as well as fully specied lexical units like
I want it. When speakers construct novel expressions, they normally use the
most specic (i.e. largest and most concrete) unit available. This is because
more specic units share more properties with the speakers conceptualisation,
and hence are easier to access in a content-addressable memory. Furthermore,
using more specic units allows the target utterance to be assembled more
easily, that is to say, in fewer steps. (Of course, speakers cannot know in
advance which method of assembling an utterance is the simplest; but if we
assume that the various possibilities are tried out in parallel, the simplest one
is the most likely to be used because it will deliver a pronounceable utterance
rst.)
This preference for specic units helps explain how speakers avoid overgen-
eralising the patterns found in their language to inappropriate contexts. The
existence of the well-entrenched unit ate pre-empts the application of the
general past-tense schema, which would produce *eated; the existence of sche-
matic units such us NP V them prevents speakers from producing expressions
such as *want they, in which the simple unit they is superimposed over the
second slot in the schema WANTER.NP want WANTED.NP; in French, the
availability of a subject + clitic pronoun + verb pattern (e.g. je le dis I say it)
pre-empts the application of the general NP V NP pattern, preventing speak-
ers from producing ungrammatical utterances such as *je dit le.
Let us now consider how one might assemble a more complex utterance,
such as this question:
(9) Where do you think they are going?
For the sake of the following discussion, we will assume that the speaker has
the following symbolic units at his or her disposal:
(10) Where AUX NONFINITE.S?

(11) WH do you think S-WH?
(12) they are going LOCATION
The rst unit, Where AUX NONFINITE.S?, is simply the lexical representa-
tion of the interrogative pronoun where. The construction represents the
speakers knowledge about how where is used in sentences, including the fact
that it requires a preposed auxiliary. Note that this fact must be explicitly rep-
resented in the grammar, since not all question words impose this requirement.
(How come, for example, does not: we say How come she is coming?, not *How
come is she coming?) Unit (11) is the prototypical LDD question construction
discussed in Chapter 9 (section 3.4); the symbol S-WH stands for a tensed
clause with a gap (i.e., an unlled slot). Unit (12) is a prefab containing a single
slot for a locative expression.
To derive the target question, the speaker must combine the three units. The
combination of (10) and (11) is a straightforward case of superimposition
with mutual elaboration:
1. where from (10) elaborates the schematic WH slot in (11);

2. do from (11) elaborates the AUX slot in (10); and
3. you think S-WH from (11) elaborates the NONFINITE.S slot in (10).
The result is this complex expression:
(13) where do you think S-WH?
The next step involves the superimposition of (12) and (13). Note that these
two components have partially incompatible specications: both contain a
location substructure, but (12) requires that the corresponding phonological
unit come after going, whereas (13) species that it should come at the begin-
ning of the sentence, before the auxiliary. To create a composite structure, the
speaker must resolve the conict, which can be done in two ways: either the
specications of (12) override those of (13), or vice versa. These two options
correspond to two dierent constructions. In ordinary questions, the WH
word is used as the framing unit and hence its requirements prevail: the result-
ing expression has interrogative word order and is used as a request for infor-
mation (14). In echo questions, the verb provides the frame, resulting in a
structure with declarative word order and a semi-declarative meaning: (15)
can be used either as a request to repeat part of the preceding utterance or
(with heavy stress on where) to express surprise, but not as a true question.
(14) where do you think they are going?

(15) you think they are going where?
Note that the approach outlined above allows for alternative ways of assem-
bling the same expression. For example, the three components of where do you
think they are going? can be combined in a dierent order: the speaker could
begin by superimposing (11) and (12) to produce WH do you think they are
going LOCATION?, and then superimpose this and (10) to derive the target
utterance. And of course if they are going LOCATION is not available as a
unit, it would have to be constructed (for example, by superimposing they are
V-ing and NP go LOCATION).
As hinted in Chapter 9, it is likely that speakers also have some more general
knowledge about LDD questions, which might be captured by this more
abstract schema
(16) WH AUX MENTAL.V (NP) S-WH?
Here, MENTAL.V stands for an untensed verb designating a mental state or

a communication event. When assembling questions such as (9), the use of this
schema would normally be pre-empted by the more specic WH do you think
S-WH? schema, but the more abstract unit can be used to produce non-
prototypical LDD questions such as where will my brother think they are going?
Alternatively, such questions could be produced by substituting will for do and
my brother for you in (11): such use would involve extension rather than instan-
tiation of the motivating schema.
Schemas (11) and (16) also sanction questions involving dependencies
reaching over more than one clause, such as examples (17) and (18):
(17) What did you say Claire thought he did?

(18) Where do you think Ben knew they were going?
Such sentences, like the more prototypical LDD question discussed earlier,
consist of a WH word followed by an auxiliary and an untensed clause with a
gap, but in this case the missing element logically belongs in the doubly embed-
ded clause (the clause that functions as the complement of thought in (17) and
knew in (18)). To be able to form such questions, the speaker must extend the
category S-WH to include expressions such as Claire thought he did and Ben
knew they were going.
4. A cognitive view of language acquisition
According to CG, language is a structured inventory of conventional linguis-

tic units. It follows that learning a language involves acquiring a repertoire of
linguistic units and making the appropriate connections between them. To
understand how language learners accomplish this, we must begin by consid-
ering the type of information that is available to them. The language addressed
to children even very young children consists predominantly of multi-word
utterances rather than simple expressions.3 The utterances occur in a rich and
fairly predictable context which enables the child to form some idea about
their meaning. They are also quite stereotypical: that is to say, some forms
repeatedly occur in similar contexts.
For example, let us suppose that during breakfast one day the child looks at
a milk carton and the mother says Do you want milk? and pours him some.
During the following weeks, the child hears You want milk, Dyou want milk?
or Do you want milk? while the mother is holding the milk carton and looking
at him, walking towards the fridge (which the child knows contains the milk),
and in other mealtime contexts. Eventually he learns that you want milk is used
when he wants milk and is oered it in other words, he acquires a xed
formula with a specic communicative function (Figure 10.6a).4 The childs

semantic representation is probably richer than the conventional adult
meaning of the expression (for example, it might incorporate his knowledge
that this is something that the mother frequently says at mealtimes and that it
is an oer of milk). It might also be incomplete or inaccurate in certain
respects. However, because humans are generally very good at guessing other
peoples communicative intentions (see Chapter 6, section 2.2), we can assume
that the childs representation will capture at least some relevant aspects of the
adult meaning.
Such formulaic phrases are the starting point of grammatical development.
To acquire units which will enable them to produce and understand novel
expressions, language learners must do three things:
1. segment the phonological representation, that is to say, identify chunks of

phonological material that function as units ([ju wnt
mlk] = [ju]<[wnt]
<[mIlk]);
2. establish correspondences between phonological chunks and salient
aspects of semantic structure (=[mlk], =[wnt] etc.); and
3. extract the component units of the formula, including constructional
schemas (in our example, [[]--[ ]]/[[]<[wnt]
<[]]).
We have a reasonably good idea of how children accomplish the rst of

these tasks. In principle, it is possible to identify word boundaries simply on
the basis of statistical regularities in the phonotactic structure of texts (Cairns
et al. 1997; Elman 1990). A variety of prosodic cues (intonation, rhythm,
stress) are also available, and we know that children make use of these (Cutler
and McQueen 1995; Mattys et al. 1999; Peters 1985).
The next stage establishing correspondences between chunks of semantic
and phonological structure would be fairly straightforward if semantic rep-
resentations were given a priori. The problem is that concepts are not available
a priori: dierent languages partition the same scenes in radically dierent
ways, so part of the learners task is to determine which aspects of a scene are
lexicalised in his or her language (cf. Chapter 7). This is especially true of the
meanings of relational words, which tend to vary cross-linguistically much
more than the meanings of concrete nouns. In addition, relational words tend
to have more abstract meanings (you can see, feel and taste milk, but you cant
see or taste wanting; you can point to milk but not to wanting) and are more
dicult to isolate conceptually (wanting cannot exist independently of the
individual who wants and the thing wanted).
So how can the child determine which semantic congurations are lexical-
ised in his language and perform the sound-to-meaning mapping? I would like
to suggest that the way information is organised in memory plays a critical role
in this process. Human memory is content-addressable, which means that:
Figure 10.6 The emergence of a schema. (a) An unanalysed formula. The holistic
representation is open to alternative semantic analyses and may contain irrelevant
contextual detail. (b) Analysed formula. Aspects of semantic structure have been
mapped onto chunks of phonological material. The schema is implicit in the
formula, but not independently accessible. (c) A low-level schema, or the lexical
representation of one sense of the verb want. Aspects of (b) shared with other want-
formulas have been entrenched through repeated use and acquired independent
status, making the schema available for production and interpretation of novel
expressions.
1. expressions which share aspects of semantic structure have partly overlap-

ping semantic representations;
2. expressions which sound similar (at either the segmental or the supraseg-
mental level) have partly overlapping phonological representations; and
3. expressions which share aspects of semantic structure and aspects of
phonological structure have partially overlapping representations at both
levels.
To see how this property can help in acquiring language, let us assume that
our learner, who has already learned to associate the phonological form
[ju]<[wnt]<[mlk] with a holistic representation of the milk-oering situa-
tion, acquires a second formula with a similar function, e.g. I want (the) ball.
The two formulas share some semantic substructures (they are both used in
situations when someone wants something) and the phonological chunk
[wnt], while other parts (e.g. the visual and motor images associated with ball,
the phonological chunk [bl]) are not shared. Once the learner has acquired
two (or more) formulas with partially overlapping representations, interesting
things start happening. The parts shared by several formulas are activated,
and hence entrenched, every time the learner hears (and later, produces) any
formula that contains them. In our example, the learners representations of
the phonological chunk [wnt] and semantic relationship between the experi-
encer and the desired object become active every time either formula is used;
and of course [mlk] and [bl] also occur in other expressions, not just as
objects of the verb want, so they, too, are activated in a variety of situations.
The elements making up the substructures shared by several formulas are fre-
quently co-activated, and eventually become units in their own right. In this
way, shared representation results in analysis: the internal structure of the
formula becomes visible to the learner (Figure 10.6b). In other words, the
component parts which are implicit in the unanalysed formula gradually
become more salient because certain congurations of semantic and phono-
logical material frequently co-occur. At this stage, the learner may be said to
have a mini-grammar: a grammar of the expression you want milk.
This change has several important consequences. First, as the formula is
analysed into its component units, meaning becomes more compositional. In
our example, whatever meaning You want milk might have had above and
beyond that contributed by its components is not strengthened as a result of
sharing a representation with other formulas, and gradually becomes less
salient; as a result, the meaning of the component units (notably the relational
expression) is distilled from the formula. The contextual meaning may not dis-
appear altogether: adult lexicons contain many multi-word units which are
not fully compositional. These include not only idioms, but also regular
expressions whose meanings are more specic than the sum of the meanings
of their parts, such as I told you so (I told you this would happen and you
should have listened to my advice/warning) or forget it (forget it because it is
not important). It may also survive as knowledge about the kinds of situa-
tions in which it is appropriate to use a particular expression (e.g. formal v.
informal, when speaking to a social superior, when speaking to a member of
the opposite sex, and so on) and about what we call connotation (knowledge
about the emotional associations of a particular expression).5
Another important consequence of analysis is that it is now easier for the
learner to acquire new formulas, since part of their representation is already
there. Analysis is basically a chunking process: complex congurations of sub-
structures become units in their own right. The child might have needed 50
repetitions to learn you want milk, 30 repetitions to learn I want (the) ball and
10 repetitions to learn you want yoghurt; subsequently he will be able to repro-
duce other instantiations of this lexically based pattern after just one or two
learning episodes (assuming he already knows the nominals).
Once a formula has been analysed into its component units, the learner has
acquired some knowledge about want constructions which one might describe
as an incipient schema. It is, however, a passive generalisation, not readily
available for use in production of novel expressions: it has a similar status to
the -en plural in English (which is found in only three nouns: oxen, children,
brethren) or the patterns underlying the smaller classes of irregular verbs (e.g.
the buyboughtbought pattern, which applies to six or seven verbs only).
There are good reasons for believing that such generalisations are captured in
our mental grammars, but they are not normally applied to new nouns and
verbs.
As observed earlier, high-frequency forms are more entrenched, and hence
easier to access. Since components shared by several stored units are necessar-
ily accessed more frequently than any of the units which contain them, it
follows that they will become more entrenched than the larger units, and, after
a time, they will become accessible as units in their own right (Figure 10.6c).
This has little impact on high-frequency formulas, which are easy to access
anyway, but has important consequences for other aspects of language use.
First, the learner now has an alternative method of retrieving low-frequency
formulas: they can be activated by accessing the component parts and allow-
ing them to activate the whole. Secondly, and much more importantly, the
schema can be combined with other established units in order to form novel
expressions in other words, it becomes productive.
It should be clear from the preceding discussion that the accessibility of a
component unit, and therefore the potential to use it in novel expressions, is a
matter of degree. The more accessible the schema, the easier it is to apply pro-
ductively, and hence the greater the likelihood that it will be used productively.
Schema accessibility depends on two main factors: specicity (as explained
earlier, more specic units are easier to access) and entrenchment, which is a
function of frequency (other things being equal, more frequent units are easier
to access).
However, even schemas which are generalisations over a relatively small
number of exemplars can be used productively. The irregular -en plural in

English is occasionally applied to new nouns, mostly in word-play: computer
hackers, for example, sometimes use the forms VAXen (several VAX comput-
ers), faxen, boxen, Macintoshen (Pinker 1999: 52). Even components which
occur in a single lexicalised unit are sometimes extended to new items. For
example, the pseudo-morpheme -k in fork can be used to produce novel words
like threek (meaning fork with three tines). This is sometimes found in
humorous musings about language (there are a number of examples on the
web); but it has also been spontaneously produced by a pre-school child
(Kaplan 1998). It should be stressed, however, that such uses are rare and
untypical in that they require extra eort on the part of both speaker and
addressee. Normal language use, in contrast, relies on well-established units
and is automatic and relatively eortless.
Thus, the capacity for grammatical productivity emerges gradually as a
result of analysis of rote-learned phrases and is a by-product of the way sym-
bolic units are stored in long-term memory. A schema is implicit in its instan-
tiations; and schema extraction involves nding it in the stored exemplars
and strengthening it until it becomes accessible as a unit in its own right. For
lexically specic patterns, this process is facilitated by the shared phonologi-
cal material. Higher-level schemas (e.g. a verbgeneral transitive construction)
are more dicult to nd because they have no shared phonological content
and rather abstract meanings. However, as explained above, such general
schemas have the same structure as their instantiations (i.e. low-level schemas
and lexically xed units), and therefore can be derived from lexically specic
units by ltering out irrelevant details while preserving the overall structure.
This account ts in very well with the developmental data discussed in the
last chapter: the tight correlation between vocabulary size and grammatical
development and the observed progression from highly stereotypical to more
exible usage. Since formulas are essentially big words, we would expect the
acquisition of both types of units to be mediated by the same mental mecha-
nism. More exible use requires constructional schemas, but since these
emerge from formulas with partially overlapping representations, the develop-
ment of productivity depends on the prior acquisition of a repertoire of rote-
learned phrases.
The above proposal also oers some insight into how learners discover the
meanings of relational words. Unlike most nouns, which are conceptually rel-
atively autonomous, the meanings of relational words such as verbs and prep-
ositions presuppose the entities participating in the relation: one cannot
conceptualise a relation without also conceptualising the entities involved in
it, even if the latter are conceived only in very schematic terms. This is easiest
to see with converse terms like own and belong to, above and below, in front of
and behind, before and after, give and receive, buy and sell etc. Every act of
giving is also an act of receiving: if you give me a lollipop, its also necessarily
true that I receive a lollipop from you. Therefore in order to learn the dier-
ence between give and receive children must match conceptualisations of

events with entire clauses rather than with single words, and the units they
extract must preserve information about which chunk of phonology corre-
sponds to which aspect of semantic structure. In other words, the verb give is
a linguistic unit consisting of four component units: GIVER-NP give RECIP-
IENT-NP GIFT-NP.
Learners also need to rely on information about the linguistic context to dis-
cover more subtle aspects of meaning. Take a set of near-synonyms like shake,
shiver, shudder, quake, quiver, tremble and vibrate. These seven verbs all desig-
nate a similar type of movement, but they are not fully synonymous and
cannot be used interchangeably: for instance, if you think of something ter-
rible, you might shudder or tremble but you wouldnt vibrate. How do we learn
these dierences in meaning? Clearly not from denitions given by more expe-
rienced language users, since most people nd the dierences very dicult to
verbalise. We also do not learn them by consulting a dictionary, as dictionary
denitions are typically circular: according to the Chambers English
Dictionary, for example, quake means quiver or vibrate,
vibrate means shake or tremble, and so on.
Learners can glean some hints about the meaning of such near-synonyms
by paying attention to the general context in which they are used: for instance,
the kind of situation that might make a person shake or tremble. However, the
single most useful and reliable source of information about such subtle seman-
tic dierences is the types of words that each verb collocates with. As shown
in a corpus study by Atkins and Levin (1995), the seven shaking verbs tend to
take dierent kinds of subjects: a person or a voice can shake or tremble, but
not shudder; a machine can vibrate, but it does not shiver; light or a shadow
can quiver or tremble, but they do not shake. Thus, learning the semantic
dierence between shake, shudder, tremble etc. is (to a large extent) learning
what classes of words they collocate with in other words, acquiring schemas
such as tremble, tremble, vibrate and so on.6
Note that apart from shake, all of these verbs are fairly literary, and hence are
normally acquired relatively late. This suggests that partially underspecied
phrase-sized units continue to play an important role in later language acqui-
sition as well.
5. More on regularity
The account outlined above emphasises stored units and low-level patterns
which are learned piecemeal. This seems to be at odds with a basic tenet of
modern linguistics, namely, that languages are highly systematic: in each lan-
guage, there are some broad regularities shared by many dierent elements,
characteristic of the language as a whole: word-order patterns, agreement rules,
structural parallels between dierent types of constructions (for example,
interrogatives, relatives) etc. Where do such patterns come from, and what is
their status? If each verb is a mini-grammar acquired independently of other
verbs ([[]--[]]/[[]<[hv]<[]], [[]--
[]]/ [[]<[it]<[]], [[]--[]]/[[]<[rat]<[]], etc.), why
does English have a xed word order (SVO) shared by all (transitive) verbs, and
why do all verbs have to agree with their subjects? Why dont dierent verbs
impose dierent word-order and agreement requirements?
Before we address this question, it is worth pointing out that modern lin-
guistics has tended to overemphasise regularity. There is, in fact, a great deal
of irregularity in language: very few rules are completely exceptionless. This is
most evident in morphology, where the existence of minor patterns and out-
right exceptions is quite commonplace. What is more, many languages have
classes of words to which the normal morphological rules do not apply, or
apply only optionally. It was noted in Chapter 8, for example, that some Polish
nouns are not inected for case in other words, the usual rules of grammar,
which require inected forms in certain contexts, do not apply to these nouns.
One can, of course, say that the rules do apply, but their operation is invisible:
the words are inected for whatever case is required, but the inections are
realised as zero. This, however, is regularity by at: the linguists grammar is
regular but it is not clear that the language itself is.
Similar observations apply to syntax. The English passive construction is
sometimes described as a product of fully general rules which convert an
underlying structure in which the direct object follows the verb into one with
passive word order. Not all transitive verbs in English, however, can be used
in the passive construction: cf. the unacceptability of *A book is had by her,
*You are suited by this dress, *A good life was lived by Robin. One might wish
to claim that have, suit and live are not truly transitive but this amounts to
saying that a transitive verb is a verb which passivises, which makes the claim
that the passive construction is fully general vacuous.
Word-order and agreement patterns may also be to some extent lexically
specic. For example, in French, most adjectives come after the noun, but
some precede it, and a few can occur in either position, depending on their
meaning (cf. femme ancienne old woman/wife and ancienne femme ex-wife).
Even English, with its relatively rigid word order, is not fully consistent: in
addition to the normal adjectivenoun order, we also have nounadjective in
expressions such as secretary general, court martial, proof positive, sister-
german, battle royal. Or consider British Sign Language, which has three types
of verbs: the so-called plain verbs are not marked for agreement and are
strongly associated with SVO word order; agreement verbs, as their name sug-
gests, do require agreement (with both subject and object), and normally
appear in sentence-nal position; and spatial verbs are marked for location
and movement rather than subject and object (Sutton-Spence and Woll 1999).
Thus, in this case we have three dierent word-order and agreement patterns
in the same language.
On the other hand, it is undeniable that there are many high-level regular-
ities in all human languages, and these require an explanation. As noted in
Chapter 6, grammars are to some degree shaped by external discourse pres-
sures and processing constraints: some utterances are preferred by language
users because they are easier to process or are less likely to lead to ambiguity,
and hence their use spreads. Since the same pressures act on dierent construc-
tions, the result is that individual constructions converge towards the same
points. Such functional pressures still leave considerable variety (since there is
generally more than one workable solution to a particular communication
problem), but they do reduce it since they eliminate some variants.
There are also pressures for systematicity. Systems in which there are several
competing patterns are inherently unstable, as it is virtually inevitable that
speakers will occasionally overgeneralise some of them, sometimes because of
memory failure, and sometimes to achieve a special eect. Such innovations
may be picked up by other speakers, which results in uctuation in the fre-
quency of the individual variants. However, once a particular variant gains a
clear advantage over the other(s), whether by chance or as a result of a lan-
guage fad, it will tend to spread (since, other things being equal, more
entrenched patterns are easier to access) and may eventually monopolise the
grammar. Of course, not all language change is regularising: processes which
result in the emergence of irregular structures are also very common other-
wise all mature languages would be completely regular. However, research on
the emergence of creole languages suggests that regularisation processes pre-
dominate in the early stages of the development of a language (Aitchison
1989; Mhlhusler 1986).
6. Future directions
The approach to language outlined above has much to recommend it, and has
been applied successfully to a variety of linguistic domains (see for example
Casad 1996b; Casad and Palmer 2003; Cienki et al. 2001; Foolen and van der
Leek 2000; Langacker 1991a; Rudzka-Ostyn 1988). However, Cognitive
Grammar is a relatively young endeavour, and is only beginning to address
certain issues, among them phenomena such as extraction, binding and
control, which have been the mainstays of generative linguistics. While there
is some promising work in these areas (see e.g. Deane 1993; Langacker 1995;
van Hoek 1997), it remains to be seen whether CG can provide a workable
alternative to generative accounts of these aspects of a speakers linguistic
knowledge.
Furthermore, although cognitive linguists are ocially committed to devel-
oping an account of language that is usage-based and rmly grounded in
human cognition, in practice, only a few have begun to go beyond the tradi-
tional introspective methods. It is to be hoped that more researchers will move
in this direction in the future. A rm empirical basis is indispensable for work

that purports to be psychologically realistic because, as stressed repeatedly
throughout this book, speakers knowledge about their language may be less
abstract than is commonly believed, and may dier considerably from one
individual to the next.
Another challenge for future research will be to develop CG accounts of
language acquisition and processing. The approach to language production
outlined above has been applied quite successfully to the speech of young chil-
dren (Dabrowska and Lieven submitted); it remains to be seen how well it will
work with adult data. With respect to language acquisition, there is a large
body of research either directly inspired by or compatible with CG (compre-
hensively reviewed in Tomasello 2003). Most of it has focused on early devel-
opment: how children learn words and multi-word units and extract the rst
generalisations from them. There is much less work on later language devel-
opment addressing such crucial issues as how early piecemeal knowledge is
integrated into a system (although Tomasello has some interesting proposals),
or how children learn about dierent construals of the same scene. However,
now that cognitive linguists have begun to work with representatives of other
language sciences, there are good prospects for progress in these areas in the
future.
Notes
1. For a detailed discussion of determiners and tense/agreement marking in

English from a CG perspective, see Langacker (1991a: chs 3 and 6). A more
accessible introduction can be found in Taylor (2002: chs 18 and 20).
2. Readers familiar with the CG framework will note a signicant departure
from standard Langackerian notation: while the subparts of semantic
structure representing the arguments of the verb are linked to their corre-
sponding phonological structures (for example, [] is connected to
[
kt]), there is no direct link between the semantic representation of the
verb and the corresponding phonological unit only between the whole
relational predication with its arguments and the corresponding phonolog-
ical structure. Likewise, in Figure 10.1b, there is no direct link between the
semantic notion of plurality and its phonological exponent. This notation
is meant to reect the fact that these lexical units are not single words, but
constructions: that is to say, their representations contain structures which
must be elaborated by other units.
3. Only about 78 per cent of the input are single-word utterances (Tomasello
2003), and these are predominantly interjections such as yeah, no, oh, OK
that is to say, words which do not enter into syntactic relationships with
other words.
4. For the sake of the following discussion, I assume that the unit that the
child extracts does not contain the initial auxiliary, since it is phonologi-
cally indistinct and often omitted in informal speech. It is also possible, of
course, that the child does represent the full adult form (do you want milk?)
or just want milk. This will obviously aect subsequent development: if the
child extracts the larger unit, he will have more analysis to do, and if the
original unit is want milk, he will have to learn some larger units at a later
stage.
5. These non-compositional aspects of meaning are usually regarded as an
extra that somehow gets added to the meanings contributed by the lexical
items when utterances are interpreted in context. However, from a devel-
opmental point of view, it is the other way around: the contextual meaning
is more basic, and the meanings of individual lexical items are carved out
of it.
6. Gleitman makes a very similar point in her argument for syntactic boot-
strapping (the hypothesis that children discover verb meanings by observ-
ing which syntactic frames verbs can occur in: see Gleitman 1990;
Gleitman and Gillette 1995). However, the above discussion suggests that
our knowledge about the distributional properties of words is much more
specic than a purely syntactic subcategorisation frame.
Bibliography
Abbot-Smith, K., Lieven, E. V. and Tomasello, M. (2001), What preschool children

do and do not do with ungrammatical word orders, Cognitive Development, 16,
67992.
Aitchison, J. (1989), Spaghetti junctions and recurrent routes: some preferred path-
ways in language evolution, Lingua, 77, 15171.
Akhtar, N. (1999), Acquiring basic word order: evidence for data-driven learning of
syntactic structure, Journal of Child Language, 26, 33956.
Akhtar, N. and Tomasello, M. (1997), Young childrens productivity with word order
and verb morphology, Developmental Psychology, 33, 95265.
Albert, M. and Obler, L. (1978), The Bilingual Brain: Neuropsychological and
Neurolinguistic Aspects of Bilingualism, New York: Academic Press.
Alegre, M. and Gordon, P. (1999), Frequency eects and the representational status
of regular inections, Journal of Memory and Language, 40, 4161.
Altenberg, B. (1998), On the phraseology of spoken English: the evidence of recurrent
word-combinations, in A. P. Cowie (ed.), Phraseology: Theory, Analysis and
Applications, Oxford: Clarendon Press, pp. 10122.
Altmann, G. T. M. and Kamide, Y. (1999), Incremental interpretation at verbs:
restricting the domain of subsequent reference, Cognition, 73, 24764.
Altmann, G. T. M., Garnham, A., van Nice, K. and Henstra, J. A. (1998), Late closure
in context, Journal of Memory and Language, 38, 45984.
Anderson, J. (1971), The Grammar of Case: Towards a Localistic Theory, Cambridge:
Cambridge University Press.
Anderson, J. (1994), Localism, in R. E. Asher and J. M. Y. Simpson (eds), The
Encyclopedia of Language and Linguistics (Vol. 4), Oxford: Pergamon Press, pp.
227682.
Anderson, S. R. and Lightfoot, D. W. (2002), The Language Organ: Linguistics As
Cognitive Physiology, Cambridge: Cambridge University Press.
Aram, D. M. (1998), Acquired aphasia in children, in M. T. Sarno (ed.), Acquired
Aphasia (3rd edn), San Diego: Academic Press, pp. 45180.
Atkins, B. T. S. and Levin, B. (1995), Building a corpus: a linguistic and lexicograph-
ical look at some near-synonyms, International Journal of Lexicography, 8, 85114.
Baayen, H., Dijkstra, T. and Schreuder, R. (1997), Singulars and plurals in Dutch: evi-
dence for a parallel dual-route model, Journal of Memory and Language, 37, 94117.
Baker, N. D. and Nelson, K. E. (1984), Recasting and related conversational tech-
niques for triggering syntactic advances by young children, First Language, 5, 321.
Bibliography 231
Baldo, J. V., Elder, J. T., Larsen, J., Dronkers, N. F., Redfern, B. and Ludy, C.
(2001), Is cognition intact in patients with aphasia?, Brain and Language, 79,
647.
Barsalou, L. W. (1992), Cognitive Psychology: An Overview for Cognitive Scientists,
Hillsdale, NJ: Lawrence Erlbaum.
Barsalou, L. W. (1993), Flexibility, structure and linguistic vagary in concepts: mani-
festations of a compositional system of perceptual symbols, in A. C. Collins, S. E.
Gathercole and M. A. Conway (eds), Theories of Memories, London: Lawrence
Erlbaum, pp. 29101.
Barsalou, L. W. (1999), Perceptual symbol systems, Behavioral and Brain Sciences, 22,
577609.
Barsalou, L. W., Yeh, W., Luka, B. J., Olseth, K. L., Mix, K. S. and Wu, L. (1993),
Concepts and meaning, Papers from the Parasession on Conceptual Representations,
Chicago Linguistic Society, 29, 2361.
Basso, A., Della Sala, S. and Fabrola, M. (1987), Aphasia arising from purely deep
lesions, Cortex, 23, 2944.
Bates, E. (1994), Modularity, domain specicity and the development of language,
Discussions in Neuroscience, 10, 13649.
Bates, E. (1997), On language savants and the structure of the mind, International
Journal of Bilingualism, 1, 16386.
Bates, E. (1999), Language and the infant brain, Journal of Communication Disorders,
32, 195205.
Bates, E. and Goodman, J. (1997), On the inseparability of grammar and the lexicon:
evidence from acquisition, aphasia, and real-time processing, Language and
Cognitive Processes, 12, 50784.
Bates, E. and Goodman, J. (1999), On the emergence of grammar from the lexicon,
in B. MacWhinney (ed.), The Emergence of Language, Mahwah, NJ: Lawrence
Erlbaum, pp. 2979.
Bates, E., Bretherton, I. and Snyder, L. (1988), From First Words to Grammar:
Individual Dierences and Dissociable Mechanisms, Cambridge: Cambridge
University Press.
Bates, E., Dale, P. S. and Thal, D. (1995), Individual dierences and their implications
for theories of language development, in P. Fletcher and B. MacWhinney (eds), The
Handbook of Child Language, Oxford: Blackwell, pp. 96151.
Bates, E., Wulfeck, B. and MacWhinney, B. (1991), Cross-linguistic research in
aphasia: an overview, Brain and Language, 41, 12348.
Bates, E., Harris, C., Marchman, V. and Wulfeck, B. (1995), Production of complex
syntax in normal ageing and Alzheimers disease, Language and Cognitive
Processes, 10, 487593.
Bates, E., Thal, D., Trauner, D., Fenson, J., Aram, D., Eisele, J. and Nass, R. (1997),
From rst words to grammar in children with focal brain injury, Developmental
Neuropsychology, 13, 44776.
Bavin, E. L. (1990), Locative terms and Warlpiri acquisition, Journal of Child
Language, 17, 4366.
Behrens, H. (2001), Learning multiple regularities: evidence from overgeneralization
errors in the German plural, in A. H.-J. Do, L. Dominguez and A. Johansen (eds),
Proceedings of the 26th Annual Boston University Conference on Language
Development, Somerville, MA: Cascadilla Press, pp. 7283.
Behrens, H. (submitted), How to learn a minority default: the acquisition of the

German -s plural.
Belliger, D. and Gleason, J. (1982), Sex dierences in parental directives to young chil-
dren, Journal of Sex Roles, 8, 112339.
Bellugi, U. (1971), Simplication in childrens language, in R. Huxley and E. Ingram
(eds), Language Acquisition: Models and Methods, London: Academic Press, pp.
95119.
Bellugi, U., Wang, P. P. and Jernigan, T. L. (1994), Williams syndrome: an unusual neuro-
psychological prole, in S. H. Broman and J. Grafman (eds), Atypical Cognitive
Decits in Developmental Disorders, Hillsdale, NJ: Lawrence Erlbaum, pp. 2356.
Bellugi, U., Marks, S., Bihrle, A. and Sabo, H. (1993), Dissociation between language
and cognitive functions in Williams syndrome, in D. Bishop and K. Mogford (eds),
Language Development in Exceptional Circumstances, Hove: Lawrence Erlbaum, pp.
17789.
Bellugi, U., Lichtenberger, L., Jones, W., Lai, Z. and St George, M. (2000), The neuro-
cognitive prole of Williams syndrome: a complex pattern of strengths and weak-
nesses, Journal of Cognitive Neuroscience, 12, 729.
Benedict, H. (1979), Early lexical development: comprehension and production,
Journal of Child Language, 6, 183200.
Berko, J. (1958), The childs learning of English morphology, Word, 14, 15077.
Berman, R. (1993), Marking verb transitivity in Hebrew-speaking children, Journal
of Child Language, 20, 64170.
Bickerton, D. (1996), Language and Human Behaviour, London: UCL Press.
Bierwisch, M. (1970), Semantics, in J. Lyons (ed.), New Horizons in Linguistics,
Harmondsworth: Penguin, pp. 16684.
Bierwisch, M. and Schreuder, R. (1992), From concepts to lexical items, Cognition,
42, 2360.
Bishop, D. (1993), Language development after focal brain damage, in D. Bishop and
K. Mogford (eds), Language Development in Exceptional Circumstances, Hove:
Lawrence Erlbaum, pp. 20319.
Bishop, D. V. M. (1983), Linguistic impairment after left hemidecortication for infan-
tile hemiplegia? A reappraisal, Quarterly Journal of Experimental Psychology, 35A,
199207.
Bishop, D. V. M. (1992), The underlying nature of specic language impairment,
Journal of Child Psychology and Psychiatry, 33, 366.
Bishop, D. V. M. (2002), Motor immaturity and specic speech and language impair-
ment: evidence for a common genetic basis, American Journal of Medical Genetics,
114, 5663.
Bishop, D. V. M., North, T. and Donlan, C. (1995), Genetic basis of specic language
impairment: evidence from a twin study, Developmental Medicine and Child
Neurology, 37, 5671.
Blackwell, A. and Bates, E. (1995), Inducing agrammatic proles in normals: evidence
for the selective vulnerability of morphology under cognitive resource limitation,
Journal of Cognitive Neuroscience, 7, 22857.
Bloom, L., Lifter, K. and Hatz, J. (1980), Semantics of verbs and the development
of verb inection in child language, Language, 56, 386412.
Bod, R. (1998), Beyond Grammar: An Experience-Based Theory of Language, Stanford,
CA: CSLI Publications.
Bibliography 233
Bod, R. (2001), Sentence memory: storage vs. computation of frequent sentences.

Retrieved 19 October 2003 from http://turing.wins.uva.nl/~rens/cuny2001.pdf.
Boesch, C. (1993), Aspects of transmission of tool-use in wild chimpanzees, in K. G.
Gibson and T. Ingold (eds), Tools, Language and Cognition in Human Evolution,
Cambridge: Cambridge University Press, pp. 17183.
Boesch, C. and Boesch-Acherman, H. (2000), The Chimpanzees of the Tai Forest:
Behavioural Ecology and Evolution, Oxford: Oxford University Press.
Borer, H. and Wexler, K. (1987), The maturation of syntax, in T. Roeper and E.
Williams (eds), Parameter-setting and Language Acquisition, Dordrecht: Reidel, pp.
12372.
Boudelaa, S. and Gaskell, M. G. (2002), A re-examination of the default system for
Arabic plurals, Language and Cognitive Processes, 17, 32143.
Bowerman, M. (1976), Semantic factors in the acquisition of rules for word use and
sentence construction, in D. Morehead and A. Morehead (eds), Normal and
Decient Child Language, Baltimore: University Park Press, pp. 99180.
Bowerman, M. (1993), Learning a semantic system: what role do cognitive predispo-
sitions play?, in P. Bloom (ed.), Language Acquisition: Core Readings, New York:
Harverster Wheatsheaf, pp. 32963.
Bowerman, M. (1996), The origins of childrens spatial semantic categories: cognitive
versus linguistic determinants, in J. Gumperz and S. C. Levinson (eds), Rethinking
Linguistic Relativity, Cambridge: Cambridge University Press, pp. 14576.
Braine, M. (1963), The ontogeny of English phrase structure, Language, 39, 114.
Braine, M. (1976), Childrens rst word combinations, Monographs of the Society for
Research in Child Development, 41 (Serial No. 167), 1104.
Braze, D., Shankweiler, D., Ni, W. and Palumbo, L. C. (2002), Readers eye movements
distinguish anomalies of form and content, Journal of Psycholinguistic Research,
31, 2544.
Briscoe, T. (1994), Prospects for practical parsing of unrestricted text: robust statisti-
cal parsing techniques, in N. Oostdijk and P. de Haan (eds), Corpus-Based Research
into Language, Amsterdam: Rodopi, pp. 6795.
Brooks, P. J. and Tomasello, M. (1999), Young children learn to produce passives with
nonce verbs, Developmental Psychology, 35, 2944.
Brown, C. M., Hagoort, P. and Kutas, M. (2000), Postlexical integration processes in
language comprehension: evidence from brain-imaging research, in M. S. Gazzaniga
(ed.), The New Cognitive Neurosciences, Cambridge, MA: MIT Press, pp. 88195.
Brown, J. W. (1975), On the neural organization of language: thalamic and cortical
relationships, Brain and Language, 2, 1830.
Brown, P. (1994), The INs and ONs of Tzeltal locative expressions: the semantics of
static descriptions of location, Linguistics, 32, 74390.
Brown, P. and Levinson, S. (1993), Uphill and downhill in Tzeltal, Journal of
Linguistic Anthropology, 3, 4674.
Brown, R. (1968), The development of wh-questions in child speech, Journal of
Verbal Learning and Verbal Behavior, 7, 27990.
Brown, R. (1973), A First Language: The Early Stages, Cambridge, MA: Harvard
University Press.
Brown, R. and Hanlon, C. (1970), Derivational complexity and order of acquisition
in child speech, in J. R. Hayes (ed.), Cognition and the Development of Language,
New York: Wiley, pp. 1153.
Brugman, C. (1983), The use of body-part terms as locatives in Chalcatongo Mixtec, in

A. Schlichter, W. Chafe and L. Hinton (eds), Studies in Mesoamerican Linguistics,
Berkeley: Survey of California and Other Indian Languages, Report No. 4, pp. 23590.
Brugman, C. (1988), The Story of Over: Polysemy, Semantics and the Structure of the
Lexicon, New York: Garland.
Brugman, C. and Macaulay, M. (1986), Interacting semantic systems: Mixtec expres-
sions of location, Proceedings of the Twelfth Annual Meeting of the Berkeley
Linguistics Society, 31527.
Bull, M. and Aylett, M. (1998), An analysis of the timing of turn-taking in a corpus
of goal-orientated dialogue, in Proceedings of ICSLP-98, Sidney, Australia (Vol. 4),
pp. 11758.
Bybee, J. L. (1995), Regular morphology and the lexicon, Language and Cognitive
Processes, 10, 42555.
Bybee, J. L. (2000), The phonology of the lexicon: evidence from lexical diusion, in
M. Barlow and S. Kemmer (eds), Usage-Based Models of Language, Cambridge:
Cambridge University Press, pp. 6585.
Bybee, J. L. and Scheibman, J. (1999), The eect of usage on degrees of constituency:
the reduction of dont in English, Linguistics, 37, 57596.
Bybee, J. L. and Slobin, D. I. (1982), Rules and schemas in the development and use
of the English past tense, Language, 58, 26589.
Cairns, P., Shillcock, R., Chater, N. and Levy, J. (1997), Bootstrapping word boundar-
ies: a bottom-up corpus-based approach to speech segmentation, Cognitive
Psychology, 33, 11153.
Calvin, W. H. (1983), A stones throw and its launch window: timing precision and its
implications for language and hominid brains, Journal of Theoretical Biology, 104,
12135.
Calvin, W. H. (1993), The unitary hypothesis: a common neural circuitry for novel
manipulations, language, plan-ahead, and throwing?, in K. R. Gibson and T. Ingold
(eds), Tools, Language and Cognition in Human Evolution, Cambridge: Cambridge
University Press, pp. 23050.
Calvin, W. H. and Bickerton, D. (2000), Lingua ex Machina: Reconciling Darwin and
Chomsky with the Human Brain, Cambridge, MA: MIT Press.
Calvin, W. H. and Ojemann, G. A. (1994), Conversations with Neils Brain: The Neural
Nature of Thought and Language, Reading, MA: Addison-Wesley.
Cameron-Faulkner, T., Lieven, E. and Tomasello, M. (in press), A construction-based
analysis of child-directed speech, Cognitive Science.
Capirci, O., Sabbadini, L. and Volterra, V. (1996), Language development in Williams
syndrome: a case study, Cognitive Neuropsychology, 13, 101739.
Carey, S. (1978), The child as a word learner, in M. Halle, J. Bresnan and G. A. Miller
(eds), Linguistic Theory and Psychological Reality, Cambridge, MA: MIT Press, pp.
26493.
Carpenter, M., Akhtar, N. and Tomasello, M. (1998), Fourteen through 18-month-old
infants dierentially imitate intentional and accidental actions, Infant Behavior and
Development, 21, 31530.
Casad, E. H. (1977), Location and direction in Cora discourse, Anthropological
Linguistics, 19, 21641.
Casad, E. H. (1982), Cora locationals and structured imagery. Unpublished PhD dis-
sertation, University of California, San Diego.
Bibliography 235
Casad, E. H. (1988), Conventionalization of Cora locationals, in B. Rudzka-Ostyn

(ed.), Topics in Cognitive Grammar, Amsterdam and Philadelphia: Benjamins, pp.
593645.
Casad, E. H. (1996a), What good are locationals, anyway?, in M. Ptz and R. Dirven
(eds), The Construal of Space in Language and Thought, Berlin: Mouton de Gruyter,
pp. 24967.
Casad, E. H. (1996b), Cognitive Linguistics in the Redwoods: The Expansion of a New
Paradigm in Linguistics, Berlin: Mouton de Gruyter.
Casad, E. H. and Langacker, R. (1985), Inside and outside in Cora grammar,
International Journal of American Linguistics, 51, 24781.
Casad, E. H. and Palmer, G. B. (eds) (2003), Cognitive Linguistics and Non-Indo-
European Languages, Berlin: Mouton de Gruyter.
Caselli, C., Casadio, P. and Bates, E. (1999), A comparision of the transition from rst
words to grammar in English and Italian, Journal of Child Language, 26, 69111.
Caselli, M. C., Leonard, L. B., Volterra, V. and Campagnoli, M. G. (1993), Towards
mastery of Italian morphology: a cross-sectional study, Journal of Child Language,
20, 37793.
Cazden, C. (1968), The acquisition of noun and verb inections, Child Development,
39, 43348.
Celli, M. L., Tomonaga, M., Udono, T., Teramoto, M. and Nagano, K. (2001),
Learning processes in the acquisition of a tool-using task by captive chimpanzees,
Psychologia, 44, 7081.
Chambers English Dictionary (1988), Cambridge: Chambers.
Chiat, S. (1982), If I were you and you were me: the analysis of pronouns in a pronoun-
reversing child, Journal of Child Language, 9, 35979.
Chiat, S. (1986), Personal pronouns, in P. Fletcher and M. Garman (eds), Language
Acquisition, Cambridge: Cambridge University Press, pp. 33955.
Chipere, N. (2001), Native speaker variations in syntactic competence: implications
for rst language teaching, Language Awareness, 10, 10724.
Chipere, N. (2003), Understanding Complex Sentences: Native Speaker Variations in
Syntactic Competence, Basingstoke: Palgrave.
Choi, S. and Bowerman, M. (1991), Learning to express motion events in English and
Korean: the inuence of language-specic lexicalization patterns, Cognition, 41,
83121.
Choi, S., Bowerman, M., Mandler, J. and McDonough, L. (1999), Early sensitivity to
language-specic spatial categories in English and Korean, Cognitive Development,
14, 24168.
Chomsky, N. (1964), Formal discussion of Miller and Ervin, Monographs of the
Society for Research in Child Development, 29 (Serial No. 92), 359.
Chomsky, N. (1975), Reections on Language, New York: Pantheon.
Chomsky, N. (1980a), On cognitive structures and their development: a reply to
Piaget, in M. Piattelli-Palmarini (ed.), Language and Learning: The Debate Between
Jean Piaget and Noam Chomsky, London: Routledge and Kegan Paul, pp. 3552.
Chomsky, N. (1980b), Discussion of Putnams comments, in M. Piattelli-Palmarini
(ed.), Language and Learning: The Debate Between Jean Piaget and Noam Chomsky,
Cambridge, MA: Harvard University Press, pp. 31024.
Chomsky, N. (1986), Knowledge of Language: Its Nature, Origin and Use, New York:
Praeger.
Chomsky, N. (1991), Linguistics and cognitive science: problems and mysteries, in A.

Kasher (ed.), The Chomskyan Turn, Cambridge, MA: Blackwell, pp. 2653.
Chomsky, N. (1999), On the nature, use, and acquisition of language, in W. C. Ritchie
and T. K. Bhatia (eds), Handbook of Child Language Acquisition, San Diego:
Academic Press, pp. 3354.
Chomsky, N. (2000), New Horizons in the Study of Language and Mind, Cambridge:
Chouinard, M. M. and Clark, E. V. (2003), Adult reformulations of child errors as neg-
ative evidence, Journal of Child Language, 30, 63769.
Cienki, A. J., Luka, B. J. and Smith, M. B. (eds) (2001), Conceptual and Discourse
Factors in Linguistic Structure, Stanford, CA: CSLI Publications.
Clahsen, H. (1999), Lexical entries and rules of language: a multidisciplinary study of
German inection, Behavioral and Brain Sciences, 22, 9911060.
Clahsen, H. and Almazan, M. (1998), Syntax and morphology in Williams syndrome,
Cognition, 68, 16798.
Clahsen, H., Eisenbeiss, S. and Sonnenstuhl-Hennig, I. (1997), Morphological struc-
ture and the processing of inected words, Theoretical Linguistics, 23, 20149.
Clahsen, H., Rothweiler, M., Woest, A. and Marcus, G. F. (1992), Regular and irreg-
ular inection in the acquisition of German noun plurals, Cognition, 45, 22555.
Clark, E. V. (1973a), Non-linguistic strategies and the acquisition of word meaning,
Clark, E. V. (1973b), Whats in a word? On the childs acquisition of semantics in his
rst language, in T. Moore (ed.), Cognitive Development and the Acquisition of
Language, New York: Academic Press, pp. 65110.
Clark, E. V. (1975), Knowledge, context and strategy in the acquisition of meaning,
in D. P. Dato (ed.), Developmental Psycholinguistics: Theory and Applications.
Georgetown University Roundatable on Languages and Linguistics 1975, Washington,
DC: Georgetown University Press, pp. 7798.
Clark, E. V. (1976), Universal categories: on the semantics of classiers and childrens
early word meanings, in A. Juilland (ed.), Linguistic Studies Oered to Joseph
Greenberg, Saratoga, CA: Anma Libri, pp. 44962.
Clark, E. V. (1980), Heres the top: nonlinguistic strategies in the acquisition of loca-
tive terms, Child Development, 51, 32938.
Clark, E. V. (1982), The young word maker: a case study of innovation in the childs
lexicon, in E. Wanner and L. R. Gleitman (eds), Language Acquisition: The State of
the Art, Cambridge, MA: Cambridge University Press, pp. 390425.
Clark, E. V. and Clark, H. (1977), Psychology and Language, New York: Harcourt
Brace.
Clark, H. (1973), Space, time, semantics and the child, in T. Moore (ed.), Cognitive
Development and the Acquisition of Language, New York: Academic Press, pp.
2763.
Clark, R. (1974), Performing without competence, Journal of Child Language, 1,
110.
Clark, R. (1977), Whats the use of imitation?, Journal of Child Language, 4, 34158.
Collins COBUILD English Language Dictonary (1993), London: HarperCollins.
Comrie, B. (1976), Aspect, Cambridge: Cambridge University Press.
Conway, C. M. and Christiansen, M. H. (2001), Sequential learning in non-human
primates, Trends in Cognitive Sciences, 5, 53946.
Bibliography 237
Corina, D. (1996), Sign language and the brain: apes, apraxia and aphasia, Behavioral
and Brain Sciences, 19, 6334.
Corina, D. (1998), Aphasia in users of signed languages, in P. Coppens, Y. Lebrun and
A. Basso (eds), Aphasia in Atypical Populations, Mahwah, NJ: Lawrence Erlbaum,
pp. 261309.
Cottrell, G. W. and Plunkett, K. (1994), Acquiring the mapping from meanings to
sounds, Connection Science, 6, 379412.
Crain, S. (1991), Language acquisition in the absence of experience, Behavioral and
Brain Sciences, 14, 597650.
Crain, S. and Lillo-Martin, D. (1999), An Introduction to Linguistic Theory and
Language Acquisition, Malden, MA: Blackwell.
Crain, S. and Thornton, R. (1991), Recharting the course of language acquisition:
studies in elicited production, in N. A. Krasnegor, D. Rumbargh, M. Studdert-
Kennedy and R. Schiefelbusch (eds), Biological and Behavioral Determinants of
Language Development, Hillsdale, NJ: Lawrence Erlbaum, pp. 32137.
Croft, W. (2001), Radical Construction Grammar: Syntactic Theory in Typological
Perspective, Oxford: Oxford University Press.
Crosson, B., Parker, J., Albert, K., Warren, R., Kepes, J. and Tully, R. (1986), A case
of thalamic aphasia with postmortem verication, Brain and Language, 29, 30114.
Culicover, P. W. (1999), Syntactic Nuts: Hard Cases, Syntactic Theory and Language
Acquisition, Oxford: Oxford University Press.
Culp, R., Watkins, R., Lawrence, H., Letts, D., Kelly, D. and Rice, M. (1991),
Maltreated childrens language and speech development: abused, neglected and
abused and neglected, First Language, 11, 37790.
Curtiss, S. and de Bode, S. (1998), Linguistic outcomes for hemispherectomized chil-
dren, in A. Greenhill, M. Hughes, H. Littleeld and H. Walsh (eds), Proceedings of
the Twenty-Second Boston University Conference on Language Development, Boston:
Cascadilla Press, pp. 12133.
Curtiss, S. and Tallal, P. (1991), On the nature of the impairment in language-impaired
children, in J. F. Miller (ed.), Research on Child Language Disorders: A Decade of
Progress, Austin, TX: Pro-ed, pp. 189210.
Curtiss, S., Katz, W. and Tallal, P. (1992), Delay vs. deviance in the language acquisi-
tion of language impaired children, Journal of Speech and Hearing Research, 35,
37283.
Cutler, A. and McQueen, J. M. (1995), The recognition of lexical units in speech, in
B. de Gelder and J. Morais (eds), Speech and Reading: A Comparative Approach,
Hove: Lawrence Erlbaum, pp. 3347.
Dabrowska, E. (1993), O jezykowej idealizacji swiata, Bulletin de la Socit Polonaise
de Linguistique, XLIX, 3542.
Dabrowska, E. (1996), The linguistic structuring of events: a study of Polish perfec-
tivizing prexes, in R. Dirven and M. Ptz (eds), The Construal of Space in
Language and Thought, Berlin: Mouton de Gruyter, pp. 46790.
Dabrowska, E. (1997), The LAD goes to school: a cautionary tale for nativists,
Dabrowska, E. (2000), From formula to schema: the acquisition of English ques-
tions, Cognitive Linguistics, 11, 120.
Dabrowska, E. (2001), Learning a morphological system without a default: the Polish
genitive, Journal of Child Language, 28, 54574.
Dabrowska, E. (2004), Rules or schemas? Evidence from Polish, Language and

Cognitive Processes, 19.
Dabrowska, E. and Lieven, E. (2003), Learning the linguistic conventions: early Polish
and English syntactic combination. Paper presented at the Biennial Meeting of the
Society for Research on Child Development, Tampa, Florida.
Dabrowska, E. and Lieven, E. (submitted), Developing question constructions: lexical
specicity and usage-based operations.
Dabrowska, E. and Szczerbinski, M. (in preparation), Polish childrens productivity
with case marking: the role of regularity, type frequency, and phonological coher-
ence.
Dale, P. S. and Crain-Thoreson, C. (1993), Pronoun reversals: who, when, and why?,
Daugherty, K. and Seidenberg, M. (1992), Rules or connections? The past tense revis-
ited, Proceedings of the 14th Annual Meeting of the Cognitive Science Society,
Hillsdale, NJ: Lawrence Erlbaum, pp. 25964.
Dawkins, R. (1986), The Blind Watchmaker, Harlow: Longman.
de Len, L. (1991), Strategies in the Acquisition of Spatial Orientation in Tzotzil (draft
paper). Nijmegen: Cognitive Anthropology Research Group, Max Planck Institute
for Psycholinguistics.
de Len, L. (1994), Exploration in the acquisition of geocentric location by Tzotzil
children, Linguistics, 32, 85784.
de Villiers, J. and de Villiers, P. (1973), A cross-sectional study of the acquisition of gra-
mamtical morphemes in child speech, Journal of Psycholinguistic Research, 2, 33141.
de Villiers, J. G. and de Villiers, P. A. (1985), The acquisition of English, in D. I. Slobin
(ed.), The Crosslinguistic Study of Language Acquisition. Volume 1: The Data,
Deacon, T. (1997), The Sympolic Species: The Co-evolution of Language and the
Human Brain, London: Penguin.
Deacon, T. (2000), Evolutionary perspectives on language and brain plasticity,
Journal of Communication Disorders, 33, 27391.
Deacon, T., Schumacher, J., Dinsmore, J., Thomas, C., Palmer, P., Kott, S., Edge, A.,
Penney, D., Kassissieh, S., Dempsey, P. and Isacson, O. (1997), Histological evi-
dence of fetal pig neural cell survival after transplantation into a patient with
Parkinsons disease, Nature and Medicine, 3, 3503.
Deane, P. (1993), Grammar in Mind and Brain: Explorations in Cognitive Syntax, Berlin
and New York: Mouton de Gruyter.
Della Corte, M., Benedict, H. and Klein, D. (1983), The relationship of pragmatic
dimensions of mothers speech to the referentialexpressive distinction, Journal of
Child Language, 10, 3543.
Dmonet, J.-F., Wise, R. and Frackowiak, R. (1993), Language functions explored in
normal subjects by positron emission tomography: a critical review, Human Brain
Mapping, 1, 3947.
Dmonet, J.-F., Puel, M., Celsis, P. and Cardebat, D. (1991), Subcortical aphasia:
some proposed pathophysiological mechanisms and their rCBF correlates revealed
by SPECT, Journal of Neurolinguistics, 6, 31944.
Dmonet, J.-F., Viallard, G., Marc-Vergnes, J. P. and Rascol, A. (1992), Cerebral blood
ow correlates of word monitoring in sentences: inuence of semantic incoherence.
A SPECT study in normals, Neuropsychologia, 30, 111.
Bibliography 239
Dennis, M. (1980), Capacity and strategy for syntactic comprehension after left of
right hemidecortication, Brain and Language, 10, 287317.
Dennis, M. and Kohn, B. (1975), Comprehension of syntax in infantile hemiplegics
after cerebral hemidecortication, Brain and Language, 2, 47586.
Dennis, M. and Whitaker, H. (1976), Language acquisition following hemidecortica-
tion: linguistic superiority of the left over the right hemisphere, Brain and Language,
3, 40433.
Dick, F., Bates, E., Utman, J. A., Wulfeck, B., Dronkers, N. and Gernsbacher, M. A.
(2001), Language decits, localization, and grammar: evidence for a distributive
model of language breakdown in aphasics and normals, Psychological Review, 108,
75988.
Dickinson, D. D. (1984), First impressions: childrens knowledge of words gained
from a single exposure, Applied Psycholinguistics, 5, 35973.
Diessel, H. and Tomasello, M. (2000), The development of relative clauses in sponta-
neous child speech, Cognitive Linguistics, 11, 13151.
Diessel, H. and Tomasello, M. (2001), The acquisition of nite complement clauses in
English: a corpus-based analysis, Cognitive Linguistics, 12, 97141.
Dockrell, J. and Campbell, R. (1986), Lexical acquisition strategies in the preschool
child, in S. Kuczaj and M. Barrett (eds), The Development of Word Meaning, New
York: Springer-Verlag, pp. 12154.
Drews, E. (1999), Pitfalls in tracking the psychological reality of lexically based and
rule-based inection, Behavioral and Brain Sciences, 22, 10223.
Dronkers, N. N., Redfern, B. B. and Knight, R. T. (2000), The neural architecture of
language disorders, in M. S. Gazzaniga (ed.), The New Cognitive Neurosciences (2nd
edn), Cambridge, MA: MIT Press, pp. 94958.
Du Bois, J. (1987), The discourse basis of ergativity, Language, 63, 80555.
Dunlea, A. (1989), Vision and the Emergence of Meaning: Blind and Sighted Childrens
Early Language, Cambridge: Cambridge University Press.
Eberhard, K. M., Spivey-Knowleton, M. J., Sedivy, J. C. and Tanenhaus, M. K. (1995),
Eye-movements as a window into real-time spoken language comprehension in
natural contexts, Journal of Psycholinguistic Research, 24, 40934.
Ebersberger, I., Metzier, D., Schwarz, C. and Paabo, S. (2002), Genome-wide compar-
ison of DNA sequences between humans and chimpanzees, American Journal of
Human Genetics, 70, 14907.
Eeg-Olofsson, M. and Altenberg, B. (1993), Discontinuous recurrent word combina-
tions in the LondonLund corpus, in U. Fries, G. Tottie and P. Schneider (eds),
Creating and Using English Language Corpora. Papers from the Fourteenth
International Conference on English Language Research on Computerized Corpora,
Zrich, Amsterdam: Rodopi, pp. 6377.
Ellis, N. C. (2000), Frequency eects in language processing: a review with implica-
tions for theories of implicit and explicit language acquisition, Studies in Second
Language Acquisition, 24, 14388.
Elman, J. (1993), Learning and development in neural networks: the importance of
starting small, Cognition, 48, 7199.
Elman, J. L. (1990), Finding structure in time, Cognitive Science, 14, 179211.
Elman, J. L. (1999), The emergence of language: a conspiracy theory, in B.
MacWhinney (ed.), The Emergence of Language, Mahwah, NJ: Lawrence Erlbaum,
pp. 127.
Elman, J. L., Bates, E., Johnson, M. H., Karmilo-Smith, A., Parisi, D. and Plunkett,
K. (1996), Rethinking Innateness: A Connectionist Perspective on Development,
Cambridge, MA: MIT Press.
Elstner, W. (1983), Abnormalities in the verbal communication of visually-impaired
children, in A. E. Mills (ed.), Language Acquisition in the Blind Child, London:
Croom Helm, pp. 1841.
Enard, W., Przeworski, M., Fisher, S. E., Lai, C. S. L., Wiebe, V., Kitano, T., Monaco,
A. P. and Paabo, S. (2002), Molecular evolution of FOXP2, a gene involved in
speech and language, Nature, 418, 86972.
Erman, B. and Warren, B. (2000), The idiom principle and the open choice principle,
Text, 20, 2962.
Ewers, H. (1999), Schemata im mentalen Lexikon: Empirische Untersuchungen zum
Erwerb der deutschen Pluralbildung, in J. Meibauer and M. Rothweiler (eds), Das
Lexikon im Spracherwerb, Tbingen: Francke, pp. 10627.
Farrar, M. J. (1990), Discourse and the acquisition of grammatical morphemes,
Fenson, L., Dale, P. S., Resnick, J. S., Bates, E., Thal, D. and Pethick, S. (1994),
Variability in early communicative development, Monographs of the Society for
Research in Child Development, 59 (Serial No. 242).
Ferreira, F. (2003), The misinterpretation of noncanonical sentences, Cognitive
Psychology, 47, 164203.
Ferreira, F. and Clifton, C., Jr (1986), The independence of syntactic processing,
Journal of Memory and Child Language, 25, 34868.
Fiez, J. A., Tallal, P., Raichle, M. E., Miezin, F. M., Katz, W. F. and Petersen, S.
E. (1995), PET studies of auditory and phonological processing: eects of stim-
ulus characteristics and task demands, Journal of Cognitive Neuroscience, 7,
35775.
Fillenbaum, S. (1974), Pragmatic normalization: further results for some conjunctive
and disjunctive sentences, Journal of Experimental Psychology, 102, 5748.
Fillmore, C. (1978), On the organization of semantic information in the lexicon,
Papers from the Parasession on the Lexicon, Chicago Linguistic Society, 111.
Fillmore, C., Kay, P., Michaelis, L. A. and Sag, I. A. (2003), Construction Grammar,
Stanford, CA: CSLI Publications.
Fillmore, C. J. (1975), An alternative to checklist theories of meaning, Proceedings of
the First Annual Meeting of the Berkeley Linguistic Society, 12331.
Fillmore, C. J. and Kay, P. (1993), Construction grammar. Unpublished manuscript.
Fisher, S. E., Lai, C. S. L. and Monaco, A. P. (2003), Deciphering the genetic basis of
speech and language disorders, Annual Review of Neuroscience, 26, 5780.
Fletcher, S., Rowland, C. and Jackson, K. (2003), Why modelling does matter: the use
of elicitation to study wh-question production in children. Paper presented at the
Child Language Seminar, Newcastle, 911 July.
Fodor, J. A. (1970), Three reasons for not deriving kill from cause to die,
Linguistic Inquiry, 1, 42938.
Fodor, J. A. (1983), The Modularity of Mind, Cambridge, MA: MIT Press.
Fodor, J. D., Fodor, J. A. and Garrett, M. (1975), The unreality of semantic represen-
tations, Linguistic Inquiry, 4, 51531.
Fodor, J. A., Garrett, M., Walker, E. and Parkes, C. (1980), Against denitions,
Bibliography 241
Foley, W. A. and Van Valin, R. D. (1984), Functional Syntax and Universal Grammar,
Cambridge: Cambridge University Press.
Foolen, A. and van der Leek, F. (2000), Constructions in Cognitive Linguistics: Selected
Papers from the Fifth International Cognitive Linguistics Conference, Amsterdam
1997, Amsterdam: John Benjamins.
Fox, P. T., Petersen, S., Posner, M. and Raichle, M. E. (1988), Is Brocas area language
specic?, Neurology, 38 (suppl.), 172.
Fraiberg, S. (1977), The acquisition of language, in S. Fraiberg (ed.), Insights from the
Blind, London: Condor Book Souvenir Press, pp. 22147.
Frazier, L. (1987), Sentence processing: a tutorial review, in M. Coltheart (ed.),
Attention and Performance XII: The Psychology of Reading, Hove: Lawrence
Erlbaum, pp. 55986.
Friederici, A. D. (1995), The time course of syntactic activation during language pro-
cessing: a model based on neuropsychological and neurophysiological data, Brain
and Language, 50, 25981.
Gardner, H. (1974), The Shattered Mind, New York: Vintage.
Garnsey, S. M., Pearlmutter, N. J., Myers, E. and Lotocky, M. A. (1997), The contri-
bution of verb bias and plausibility to the comprehension of temporarily amgigu-
ous sentences, Journal of Memory and Language, 37, 5893.
Gasser, M. and Smith, L. B. (1998), Learning nouns and adjectives: a connectionist
account, Language and Cognitive Processes, 13, 269306.
Gathercole, S. E. and Baddeley, A. D. (1990), Phonological memory decits in lan-
guage disordered children: is there a causal relationship?, Journal of Memory and
Language, 29, 33660.
Gawlitzek-Maiwald, I. (1994), How do children cope with variation in the input? The
case of German plurals and compounding, in R. Tracy and E. Lattey (eds), How
Tolerant is Universal Grammar? Essays on Language Variability and Language
Variation, Tbingen: Niemeyer, pp. 22566.
Geeraerts, D. (1988), Where does prototypicality come from?, in B. Rudzka-
Ostyn (ed.), Topics in Cognitive Linguistics, Amsterdam: John Benjamins, pp.
20729.
Gibbs, R. W. (1980), Spilling the beans on understanding and memory for idioms in
conversation, Memory and Cognition, 8, 14956.
Gibbs, R. W. and Colston, H. L. (1995), The cognitive psychological reality of image
schemas and their transformations, Cognitive Linguistics, 6, 34778.
Gibbs, R. W., Jr and Gonzales, G. P. (1985), Syntactic frozenness in processing and
remembering idioms, Cognition, 20, 24359.
Givn, T. (1989), Mind, Code and Context: Essays in Pragmatics, Hillsdale, NJ:
Lawrence Erlbaum.
Gleitman, L. R. (1981), Maturational determinants of language growth, Cognition,
10, 10314.
Gleitman, L. R. (1990), The structural sources of verb meanings, Language
Acquisition, 1, 355.
Gleitman, L. R. and Gillette, J. (1995), The role of syntax in verb learning, in P.
Fletcher and B. MacWhinney (eds), The Handbook of Child Language, Oxford:
Blackwell, pp. 41327.
Goldberg, A. E. (1995), Constructions: A Construction Grammar Approach to
Argument Structure, Chicago: University of Chicago Press.
Goldeld, B. A. and Reznick, J. S. (1990), Early lexical acquisition: rate, content and
vocabulary spurt, Journal of Child Language, 17, 17184.
Goldman-Eisler, F. (1968), Psycholinguistics: Experiments in Spontaneous Speech,
London: Academic Press.
Goodall, J. (1986), The Chimpanzees of Gombe: Patterns of Behavior, Cambridge, MA:
Harvard University Press.
Goodglass, H. (2001), The Assessment of Aphasia and Related Disorders, Philadelphia:
Lippincott, Williams and Williams.
Goodglass, H. and Kaplan, E. (1972), The Assessment of Aphasia and Related
Disorders, Philadelphia: Lea and Febiger.
Gordon, P. and Chafetz, J. (1990), Verb-based versus class-based accounts of action-
ality eects in childrens comprehension of passives, Cognition, 36, 22754.
Graham, M. D. (1968), Multiply Impaired Blind Children: A National Problem, New
York: American Foundation for the Blind.
Grant, J., Valian, V. and Karmilo-Smith, A. (2002), A study of relative clauses in
Williams syndrome, Journal of Child Language, 29, 40316.
Grodzinsky, Y. and Finkel, L. (1998), The neurology of empty categories, Journal of
Cognitive Neuroscience, 10, 28192.
Grosjean, F. (1980), Spoken word recognition and the gating paradigm, Perception
and Psychophysics, 28, 299310.
Grossman, M. (1980), A central processor for hierarchically structured material: evi-
dence from Brocas aphasia, Neuropsychologia, 18, 299309.
Gullo, D. F. (1981), Social class dierences in preschool childrens comprehension of
wh-questions, Child Development, 52, 73640.
Haiman, J. (1985), Iconicity in Syntax, Cambridge: Cambridge University Press.
Hall, W. S., Nagy, W. E. and Linn, R. (1984), Spoken Words: Eects of Situation
and Social Group on Oral Word Usage and Frequency, Hillsdale, NJ: Lawrence
Erlbaum.
Hampson, J. and Nelson, K. (1993), The relation of maternal language to variation in
rate and style of language acquisition, Journal of Child Language, 20, 31342.
Hamsher, K. (1998), Intelligence and aphasia, in M. T. Sarno (edn), Acquired Aphasia
(3rd edn), San Diego: Academic Press, pp. 34173.
Hare, M. and Elman, J. L. (1995), Learning and morphological change, Cognition,
56, 6198.
Hare, M., Elman, J. L. and Daugherty, K. G. (1995), Default generalization in con-
nectionist networks, Language and Cognitive Processes, 10, 60130.
Hart, B. and Risley, R. R. (1995), Meaningful Dierences in the Everyday Experience
of Young American Children, Baltimore: Paul Brooks.
Hart, B. and Risley, R. R. (1999), The Social World of Children Learning to Talk,
Baltimore: Paul Brookes.
Hauser, M. D., Chomsky, N. and Fitch, W. T. (2002), The faculty of language: what is
it, who has it, and how did it evolve, Science, 298, 156979.
Hawkins, J. A. (1994), A Performance Theory of Order and Constituency, Cambridge:
Heibeck, T. H. and Markman, E. M. (1987), Word learning in children: an examina-
tion of fast mapping, Child Development, 58, 102134.
Heine, B. (1997), Cognitive Foundations of Grammar, New York and Oxford: Oxford
University Press.
Bibliography 243
Herskovits, A. (1985), Semantics and pragmatics of locative expressions, Cognitive

Science, 9, 34178.
Herskovits, A. (1986), Language and Spatial Cognition, Cambridge: Cambridge
University Press.
Hickey, T. (1993), Identifying formulas in rst language acquisition, Journal of Child
Language, 20, 2741.
Hill, E. L. (2001), Non-specic nature of specic language impairment: a review of
the literature with regard to concomitant motor impairments, International Journal
of Language & Communication Disorders, 36, 14971.
Hirsh-Pasek, K. and Golinko, R. M. (1996), The Origins of Grammar: Evidence from
Early Language Comprehension, Cambridge, MA: MIT Press.
Hjelmslev, L. (1935), La catgorie des cas, Copenhagen: Munksgaard.
How many genes are in the human genome? (2003). Retrieved 20 September
2003 from Human Genome Project Information website, http://www.ornl.gov/
TechResources/Human_Genome/faq/genenumber.html.
Hu, Y.-H., Qiou, Y.-G. and Zhong, G.-Q. (1990), Crossed aphasia in Chinese: a clini-
cal survey, Brain and Language, 39, 34756.
Huggins, A. W. F. (1964), Distortion of the temporal pattern of speech: interruption
and alternation, Journal of the Acoustical Society of America, 36, 105564.
Huttenlocher, J. (1998), Language input and language growth, Preventive Medicine,
27, 1959.
Huttenlocher, J., Smiley, P. and Ratner, H. (1983), What do word meanings reveal
about conceptual development?, in T. B. Seiler and W. Wannenmacher (eds),
Concept Development and the Development of Word Meaning, Berlin: Springer
Verlag, pp. 21033.
Huttenlocher, J., Vasilyeva, M., Cymerman, E. and Levine, S. (2002), Language input
and child syntax, Cognitive Psychology, 45, 33774.
Huttenlocher, J., Haight, W., Bryk, A., Seltzer, M. and Lyons, T. (1991), Early vocab-
ulary growth: relation to language input and gender, Developmental Psychology, 27,
23648.
Indefrey, P. (1999), Some problems with the lexical status of nondefault inection,
Behavioral and Brain Sciences, 22, 1025.
Ingram, D. and Tyack, D. (1979), The inversion of subject NP and Aux in childrens
questions, Journal of Psycholinguistic Research, 4, 33341.
Iwamura, S. G. (1979), I don wanna dont throw: speech formulas in rst language
acquisition. Unpublished manuscript, University of Hawaii.
Jackendo, R. (1983), Semantics and Cognition, Cambridge, MA: MIT Press.
Jackendo, R. (1990), Semantic Structure, Cambridge, MA: MIT Press.
Jackendo, R. (2002), Foundations of Language: Brain, Meaning, Grammar, Evolution,
Oxford: Oxford University Press.
Jaeger, J. J., Lockwood, A. H., Kemmerer, D. L., Van Valin, R. D. and Murphy, B. W.
(1996), Positron emission tomographic study of regular and irregular verb
morphology in English, Language, 72, 45197.
Jkel, O. (1995), The metaphorical concept of mind: mental activity is manipula-
tion , in J. R. Taylor and R. E. MacLaury (eds), Language and the Cognitive
Construal of the World, Berlin: Mouton de Gruyter, pp. 197229.
Janda, L. A. (1996), Figure, ground and animacy in Slavic declension, Slavic and East
European Journal, 40, 32555.
Johnson, M. (1987), The Body in the Mind: The Bodily Basis of Reason and
Imagination, Chicago: University of Chicago Press.
Johnson, M. H. (1993), Constraints on cortical plasticity, in M. H. Johnson (ed.),
Brain Development and Cognition: A Reader, Cambridge: Blackwell, pp. 70322.
Johnson, M. H. (1998), The neural basis of cognitive development, in D. Kuhn and
R. S. Siegler (eds), Handbook of Child Psychology. Vol. 2: Cognition, Perception and
Language, New York: John Wiley and Sons, pp. 150.
Johnston, J. R. (1997), Specic language impairment, cognition and the biological
basis of language, in M. Gopnik (ed.), The Inheritance and Innateness of Grammars,
New York: Oxford University Press, pp. 16180.
Johnston, J. R. and Slobin, D. I. (1979), The development of locative expressions in
English, Italian, Serbo-Croatian and Turkish, Journal of Child Language, 6, 52945.
Jones, D., von Stienbrugge, W. and Chicralis, K. (1994), Underlying decits in
language-disordered children with central auditory processing diculties, Applied
Psycholinguistics, 15, 31128.
Jones, M. J. (1996), A longitudinal and methodological investigation of auxiliary verb
development. Unpublished thesis, University of Manchester.
Jones, S. S. and Smith, L. B. (1993), The place of perception in childrens concepts,
Cognitive Development, 8, 11339.
Jou, J. and Harris, R. J. (1991), Processing inections: dynamic processes in sentence
comprehension, Journal of Experimental Psychology: Learning Memory and
Kaplan, J. (1998), Message posted to Funknet, Funknet Archive, 13 October 1998.
Retrieved 8 December 2003 from https://mailman.rice.edu/pipermail/funknet/1998-
October.txt.
Karmilo-Smith, A. (1998), Is atypical development necessarily a window on the
normal mind/brain? The case of Williams syndrome, Developmental Science, 1,
2737.
Karmilo-Smith, A., Grant, J., Bethoud, I., Davies, M., Howlin, P. and Udwin, O.
(1997), Language and Williams syndrome: how intact is intact?, Child
Development, 68, 24662.
Karmilo-Smith, A., Tyler, L. K., Voice, K., Sims, K., Udwin, O., Howlins, P. and
Davies, M. (1998), Linguistic dissociations in Williams syndrome: evaluating recep-
tive syntax in on-line and o-line tasks, Neuropsychologia, 36, 34351.
Katz, J. (1972), Semantic Theory, New York: Harper and Row.
Katz, J. (1987), Common sense in semantics, in E. Lepore (ed.), New Directions in
Semantics, London: Academic Press, pp. 157234.
Katz, J. J. and Fodor, J. A. (1963), The structure of semantic theory, Language, 39,
170210.
Kegl, J., Senghas, A. and Coppola, M. (1999), Creation through contact: sign lan-
guage emergence and sign language change in Nicaragua, in M. DeGraf (ed.),
Language Creation and Language Change: Creolization, Diachrony and Development,
Cambridge, MA: MIT Press, pp. 179238.
Keil, K. and Kaszniak, A. W. (2002), Examining executive function in individuals with
brain injury: a review, Aphasiology, 16, 30535.
Kertesz, A. and Osmn-Sgi, J. (2001), Manifestations of aphasic symptoms in
Hungarian, Journal of Neurolinguistics, 14, 31319.
Kidd, E., Bavin, E. L. and Rhodes, B. (2001), Two-year-olds knowledge of verbs and
Bibliography 245
argument structure, in M. Almgren, A. Barrea, M.-J. Ezeizabarrena, I. Idiazabal

and B. MacWhinney (eds), Research on Child Language Acquisition. Proceedings of
the 8th Conference of the International Association for the Study of Child Language,
Somerville: Cascadilla Press, pp. 136882.
Kim, J. J., Pinker, S., Prince, A. S. and Prasada, S. (1991), Why no mere mortal has
ever own out to center eld, Cognitive Science, 15, 173218.
Kim, J. J., Marcus, G. F., Pinker, S., Hollander, M. and Coppola, M. (1994), Sensitivity
of childrens inections to grammatical structure, Journal of Child Language, 21,
173209.
Kimura, D. (1993), Neuromotor Mechanisms in Human Communication, Oxford:
Oxford University Press.
Kiparsky, P. (1985), Some consequences of lexical phonology, Phonology Yearbook,
2, 85138.
Kirby, S. (1998), Fitness and the selective adaptation of language, in J. R. Hurford,
M. Studdert-Kennedy and C. Knight (eds), Approaches to the Evolution of
Language: Social and Cognitive Bases, Cambridge: Cambridge University Press, pp.
35983.
Klima, E. S. and Bellugi, U. (1966), Syntactic regularities in the speech of children,
in J. Lyons and R. J. Wales (eds), Psycholinguistics Papers. Proceedings of the 1966
Edinburgh Conference, Edinburgh: Edinburgh University Press, pp. 183208.
Kluender, K., Diehl, R. and Killeen, P. (1987), Japanese quail can learn phonetic cat-
egories, Science, 237, 11957.
Kluender, R. (1998), On the distinction between strong and weak islands: a process-
ing perspective, in P. W. Culicover and L. McNally (eds), Syntax and Semantics 29:
The Limits of Syntax, San Diego: Academic Press, pp. 24179.
Kluender, R. and Kutas, M. (1993), Subjacency as a processing phenomenon,
Language and Cognitive Processes, 8, 573633.
Kpcke, K.-M. (1988), Schemas in German plural formation, Lingua, 74, 30335.
Kpcke, K.-M. (1998), The acquisition of plural marking in English and German
revisited: schemata versus rules, Journal of Child Language, 25, 293319.
Kottum, S. E. (1981), The genitive singular form of masculine nouns in Polish,
Scando-Slavica, 27, 17986.
Kvecses, Z. (1986), Metaphors of Anger, Pride and Love, Amsterdam: Benjamins.
Krashen, S. D. and Scarcella, R. (1978), On routines and patterns in language acqui-
sition and performance, Language Learning, 28, 283300.
Kuczaj, S. and Brannick, N. (1979), Childrens use of the wh-question modal auxiliary
placement rule, Journal of Experimental Child Psychology, 28, 4367.
Kuczaj, S. A. (1976), -ing, -s and -ed: a study of the acquisition of certain verb
inections. Unpublished PhD dissertation, University of Minnesota.
Kuczaj, S. A. and Maratsos, M. P. (1983), Initial verbs in yes-no questions: a dierent
kind of general grammatical category?, Developmental Psychology, 19, 4404.
Kuhl, P. K. (1981), Discrimination of speech by nonhuman animals: basic auditory
sensitivities conducive to the perception of speech-sound categories, Journal of the
Acoustical Society of America, 70, 3409.
Kuiper, K. (1996), Smooth Talkers: The Linguistic Performance of Auctioneers and
Sportscasters, Mahwah, NJ: Lawrence Erlbaum.
Kurcz, I., Lewicki, A., Sambor, J., Szafran, K. and Woronczak, J. (1990), Sl-ownik frek-
wencyjny polszczyzny wspl-czesnej, Krakw: Polska Akademia Nauk.
Kutas, M. and King, J. W. (1995), The potentials for basic sentence processing: dier-
entiating integrative processes, in T. Inui and J. L. McClelland (eds), Attention and
Performance XVI: Information Integration in Perception and Communication,
Labov, W. and Labov, T. (1978), Learning the syntax of questions, in R. N. Campbell
and P. T. Smith (eds), Recent Advances in the Psychology of Langauge (Vol. 4B), New
York: Plenum Press, pp. 144.
Lachter, J. and Bever, T. (1988), The relation between linguistic structure and associa-
tive theories of language learning: a constructive critique of some connectionist
learning models, Cognition, 28, 195247.
Lai, C. S. L., Fisher, S. E., Hurst, J. A., Vargha-Khadem, F. and Monaco, A. P. (2001),
A forkhead-domain gene is mutated in a severe speech and language disorder,
Nature, 413, 51923.
Lako, G. (1987), Women, Fire and Dangerous Things: What Categories Reveal about
the Mind, Chicago: Chicago University Press.
Lako, G. and Johnson, M. (1980), Metaphors We Live By, Chicago: Chicago
University Press.
Landau, B. and Gleitman, L. R. (1985), Language and Experience: Evidence from the
Blind Child, Cambridge, MA: Harvard University Press.
Langacker, R. W. (1987), Foundations of Cognitive Grammar. Vol. 1: Theoretical
Prerequisites, Stanford, CA: Stanford University Press.
Langacker, R. W. (1991a), Foundations of Cognitive Grammar. Vol. 2: Descriptive
Application, Stanford, CA: Stanford University Press.
Langacker, R. W. (1991b), Concept, Image and Symbol: The Cognitive Basis of
Grammar, Berlin: Mouton de Gruyter.
Langacker, R. W. (1995), Raising and transparency, Language, 71, 162.
Langacker, R. W. (2000), A dynamic usage-based model, in M. Barlow and S. Kemmer
(eds), Usage-Based Models of Language, Stanford, CA: CSLI Publications, pp. 163.
Lebrun, Y. and Hoops, R. (1974), Intelligence and Aphasia, Amsterdam: Swets and
Zeitlinger B.V.
Levine, S. C. and Carey, S. (1982), Up front: the acquisition of a concept and a word,
Levinson, S. C. (1994), Vision, shape and linguistic description: Tzeltal body-part ter-
minology and object description, Linguistics, 94, 791855.
Lieberman, P. (1991), Uniquely Human: The Evolution of Speech, Thought, and Seless
Behavior, Cambridge, MA: Harvard University Press.
Lieberman, P. (1998), Eve Spoke: Human Language and Human Evolution, London:
Picador.
Lieberman, P. (2002), On the nature and evolution of the neural bases of human lan-
guage, Yearbook of Physical Anthropology, 45, 3662.
Lieberman, P., Friedman, J. and Feldman, L. S. (1990), Syntax comprehension decits
in Parkinsons disease, Journal of Nervous and Mental Disease, 178, 3605.
Lieberman, P., Kako, E. T., Friedman, J., Tajchman, G., Feldman, L. S. and Jiminez,
E. B. (1992), Speech production, syntax comprehension, and cognitive decits in
Parkinsons disease, Brain and Language, 43, 16989.
Lieven, E. V. (1994), Crosslinguistic and crosscultural aspects of language addressed
to children, in C. Gallaway and B. J. Richards (eds), Input and Interaction in
Language Acquisition, Cambridge: Cambridge University Press, pp. 5673.
Bibliography 247
Lieven, E. V. (1997), Variation in a crosslinguistic context, in D. I. Slobin (ed.), The

Crosslinguistic Study of Language Acquisition. Vol. 5: Expanding the Contexts,
Mahwah, NJ: Lawrence Erlbaum, pp. 199263.
Lieven, E. V. (2003), Memory, productivity and constructions in early English multi-
word utterances. Paper presented at the Biennial Meeting of the Society for
Research on Child Development, Tampa, Florida.
Lieven, E. V., Pine, J. M. and Baldwin, G. (1997), Lexically-based learning and early
grammatical development, Journal of Child Language, 24, 187219.
Lieven, E. V., Pine, J. M. and Barnes, H. D. (1992), Individual dierences in early
vocabulary development: redening the referentialexpressive distinction, Journal
of Child Language, 19, 287310.
Lightfoot, D. (1989), The childs trigger experience: degree-0 learnability, Behavioral
Linebarger, M. C. (1989), Neuropsychological evidence for linguistic modularity, in
G. N. Carlson and M. K. Tanenhaus (eds), Linguistic Structure in Language
Processing, Dordrecht: Kluwer Academic, pp. 197238.
Linebarger, M. C. (1990), Neuropsychology of sentence parsing, in A. Caramazza
(ed.), Cognitive Neuropsychology and Neurolinguistics, Hillsdale, NJ: Lawrence
Erlbaum, pp. 55122.
Linebarger, M. C., Schwartz, M. F. and Saran, E. M. (1983), Sensitivity to grammat-
ical structure in so-called agrammatic aphasics, Cognition, 13, 36192.
Lock, A. and Colombo, M. (1996), Cognitive abilities in a comparative perspective,
in A. Lock and C. R. Peters (eds), Handbook of Human Symbolic Evolution, Oxford:
Clarendon Press, pp. 596643.
Lock, A. and Peters, C. R. (1996), Editorial appendix II: evolution of the human vocal
apparatus, in A. Lock and C. R. Peters (eds), Handbook of Human Symbolic
Evolution, Oxford: Clarendon Press, pp. 11625.
Locke, A. and Ginsborg, J. (submitted), Spoken language in the early years: the cog-
nitive and linguistic developemnt of three- to ve-year-old children from socio-
economically deprived backgrounds.
Locke, A., Ginsborg, J. and Peers, I. (2002), Development and disadvantage: implica-
tions for the early years and beyond, International Journal of Language and
Communication Disorders, 37, 315.
Logan, G. D. (1988), Toward an instance theory of automatization, Psychological
Review, 95, 492527.
MacDonald, M. C. (1994), Probabilistic constraints and syntactic ambiguity resolu-
tion, Language and Cognitive Processes, 9, 157201.
MacDonald, M. C. (1999), Distributional information in language comprehension,
production and acquisition: three puzzles and a moral, in B. MacWhinney (ed.),
The Emergence of Language, Mahwah, NJ: Lawrence Erlbaum, pp. 17796.
Maclay, H. and Osgood, C. E. (1959), Hesitation phenomena in spontaneous English
speech, Word, 15, 1944.
MacWhinney, B. (1985), Hungarian language acquisition as an exemplication of a
general model of grammatical development, in D. I. Slobin (ed.), The
Crosslinguistic Study of Language Acquisition. Vol. 2: Theoretical Issues, Hillsdale,
NJ: Lawrence Erlbaum, pp. 1069155.
MacWhinney, B. (1995), The CHILDES Project: Tools for Analyzing Talk, Hillsdale,
NJ: Lawrence Erlbaum.
MacWhinney, B. and Leinbach, J. (1991), Implementations are not conceptualiza-

tions: revising the verb learning model, Cognition, 29, 12157.
MacWhinney, B., Osmn-Sgi, J. and Slobin, D. I. (1991), Sentence comprehension in
aphasia in two clear case-marking languages, Brain and Language, 41, 23449.
Major, D. (1974), The Acquisition of Modal Auxiliaries in the Language of Children,
The Hague: Mouton.
Mandler, J. M. (1992), How to build a baby II: Conceptual primitives, Psychological
Review, 99, 587604.
Maratsos, M. (2000), More overregularizations after all: new data and discussion on
Marcus, Pinker, Ullmann, Hollander, Rosen & Xu, Journal of Child Language, 27,
183212.
Marchman, V. A. (1993), Constraints on plasticity in a connectionist model of the
English past tense, Journal of Cognitive Neuroscience, 5, 21534.
Marcus, G. F. (1998), Can connectionism save constructivism?, Cognition, 66, 15382.
Marcus, G. F. and Fisher, S. E. (2003), FOXP2 in focus: what can genes tell us about
speech and language?, Trends in Cognitive Science, 7, 25762.
Marcus, G. F., Brinkmann, U., Clahsen, H., Wiese, R. and Pinker, S. (1995), German
inection: the exception that proves the rule, Cognitive Psychology, 29, 189256.
Marcus, G. F., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J. and Xu, F. (1992),
Overregularization in language acquisition, Monographs of the Society for
Research in Child Development, 57 (Serial No. 22), 1180.
Markman, E. M. (1987), How children constrain the possible meanings of words, in
U. Neisser (ed.), Concepts and Conceptual Development: Ecological and Intellectual
Factors in Categorization, Cambridge: Cambridge University Press, pp. 25587.
Markman, E. M. (1990), Constraints children place on word meanings, Cognitive
Science, 14, 5777.
Markson, L. and Bloom, P. (1997), Evidence against a dedicated system for word
learning in children, Nature, 385, 81315.
Marshall, J. C. (1986), The description and interpretation of aphasic language disor-
der, Neuropsychologia, 24, 524.
Marslen-Wilson, W. and Tyler, L. K. (1987), Against modularity, in J. L. Gareld
(ed.), Modularity in Knowledge Representation and Natural-Language Processing,
Marslen-Wilson, W. D. (1973), Linguistic structure and speech shadowing at very
short latencies, Nature, 244, 5223.
Marslen-Wilson, W. D. (1987), Functional parallelism in spoken word-recognition,
Marslen-Wilson, W. D. and Tyler, L. K. (1997), Dissociating types of mental compu-
tation, Nature, 387, 5924.
Marslen-Wilson, W. D. and Tyler, L. K. (1998), Rules, representations, and the English
past tense, Trends in Cognitive Science, 2, 42835.
Martin, W. K., Church, K. and Patil, R. (1987), Preliminary analysis of a breadth-rst
parsing algorithm: theoretical and experimental results, in L. Bolc (ed.), Natural
Language Parsing Systems, Berlin: Springer Verlag, pp. 267328.
Martins, I. P. and Ferro, J. M. (1992), Recovery of acquired aphasia in children,
Aphasiology, 6, 4318.
Martins, I. P. and Ferro, J. M. (1993), Acquired childhood aphasia: a clinicoradiolog-
ical study of 11 stroke patients, Aphasiology, 7, 48995.
Bibliography 249
Massaro, D. W. (1994), Psychological aspects of speech perception: implications for

research and theory, in M. A. Gernsbacher (ed.), Handbook of Psycholinguistics,
San Diego: Academic Press, pp. 21963.
Matsuki, K. (1995), Metaphors of anger in Japanese, in J. R. Taylor and R. E.
MacLaury (eds), Language and the Cognitive Construal of the World, Berlin:
Mouton de Gruyter, pp. 13751.
Mattys, S. L., Jusczyk, P. W., Luce, P. A. and Morgan, J. L. (1999), Phonotactic and
prosodic eects on word segmentation in infants, Cognitive Psychology, 38, 46594.
Mazziotta, J. C. and Metter, J. E. (1988), Brain cerebral metabolic mapping of normal
and abnormal language and its acquisition during development, in F. Plum (ed.),
Language, Communication and the Brain, New York: Raven, pp. 24566.
Mazziota, J. C., Phelps, M. E., Carson, R. E. and Kuhl, D. E. (1982), Tomographic
mapping of human cerebral metabolism: auditory stimulation, Neurology, 32,
92137.
Meltzo, A. (1995), Understanding the intentions of others: re-enactment of intended
acts by 18-month-old children, Developmental Psychology, 24, 4706.
Menn, L. and Obler, L. K. (eds) (1990), Agrammatic Aphasia: A Cross-Language
Narrative Sourcebook, Amsterdam: Benjamins.
Merzenich, M. M., Jenkins, W. M., Johnston, P., Schreiner, C., Miller, S. L. and Tallal,
P. (1996), Temporal processing decits in language-learning impaired children
ameliorated by training, Science, 271, 7781.
Miller, G. A. (1986), Dictionaries in the mind, Language and Cognitive Processes, 1,
17185.
Miller, G. A. and Gildea, P. M. (1987), How children learn words, Scientic American,
257, 869.
Miller, G. A. and Johnson-Laird, P. N. (1976), Language and Perception, Cambridge:
Miller, J. E. (1972), Towards a generative semantic account of aspect in Russian,
Journal of Linguistics, 8, 21736.
Miller, J. E. (1974), A localist account of the dative case in Russian, in R. D. Brecht
and C. V. Chvany (eds), Slavic Transformational Syntax, Ann Arbor: University of
Michigan, pp. 24461.
Miller, J. L. and Jusczyk, P. W. (1989), Seeking the neurobiological bases of speech per-
ception, Cognition, 33, 11137.
Mills, A. (1985), The acquisition of German, in D. I. Slobin (ed.), The Cross-
Linguistic Study of Language Acquisition. Vol. 1: The Data, Hillsdale, NJ: Lawrence
Erlbaum, pp. 141254.
Mitchell, D. C. (1994), Sentence parsing, in M. A. Gernsbacher (ed.), Handbook of
Psycholinguistics, New York: Academic Press, pp. 357409.
Miyake, A., Carpenter, P. A. and Just, M. A. (1994), A capacity approach to syntactic
comprehension disorders: making normal adults perform like aphasic patients,
Cognitive Neuropsychology, 11, 671717.
Mogford, K. (1993), Language development in twins, in K. Mogford and D. Bishop
(eds), Language Development in Exceptional Circumstances, Hove: Lawrence
Erlbaum, pp. 8095.
Montgomery, J. (1995), Examination of phonological working memory in specically
language-impaired children, Applied Psycholinguistics, 16, 35578.
Moon, R. (1998), Frequencies and forms of phrasal lexemes in English, in A. P. Cowie
(ed.), Phraseology: Theory, Analysis and Applications, Oxford: Clarendon Press, pp.
79100.
Mueller, R. A. G. and Gibbs, R. W. (1987), Processing idioms with multiple meanings,
Journal of Psycholinguistic Research, 16, 6381.
Mhlhusler, P. (1986), Pidgin and Creole Linguistics, Oxford: Blackwell.
Mulford, R. (1988), First words of the blind, in M. D. Smith and J. L. Locke (eds),
The Emergent Lexicon: The Childs Development of a Linguistic Vocabulary, New
York: Academic Press, pp. 293338.
Mller, R.-A. (1992), Modularism, holism, connectionism: old conicts and new per-
spectives in aphasiology and neuropsychology, Aphasiology, 6, 44375.
Mller, R.-A. (1996), Innateness, autonomy, universality? Neurobiological approaches
to langauge, Behavioral and Brain Sciences, 19, 61175.
Murray, A. D., Johnson, J. and Peters, J. (1990), Fine-tuning of utterance length to
preverbal infants: eects on later language development, Journal of Child Language,
17, 51125.
Nagell, K., Olguin, K. and Tomasello, M. (1993), Processes of social learning in the
tool use of chimpanzees (Pan troglodytes) and human children (Homo sapiens),
Journal of Comparative Psychology, 107, 17486.
Nagrko, A. (1998), Zarys gramatyki polskiej (ze sl-owotwrstwem), Warsaw: PWN.
Naigles, L. R. (1990), Children use syntax to learn verb meaning, Journal of Child
Nelson, K. (1973), Structure and strategy in learning to talk, Monographs of the
Society for Research in Child Development, 38 (Serial No. 149), 1135.
Nelson, K. (1981), Individual dierences in language development: implications for
development and language, Developmental Psychology, 17, 17087.
Neville, H. J. and Bavelier, D. (2000), Specicity and plasticity in neurocognitive devel-
opment in humans, in M. S. Gazzaniga (ed.), The New Cognitive Neurosciences,
Neville, H. J., Mills, D. L. and Bellugi, U. (1993), Eects of altered auditory sensitiv-
ity and age of language acquisition on the development of language-relevant neural
systems: preliminary studies of Williams syndrome, in S. Broman and J. Grafman
(eds), Cognitive Decits in Developmental Disorders: Implications for Brain Function,
New, H. (1972), Ranking of constraints on /t, d/ deletion in American English: a sta-
tistical analysis, in W. Labov (ed.), Locating Language in Time and Space, New
York: Academic Press, pp. 3754.
Ni, W., Crain, S. and Shankweiler, D. (1996), Sidestepping garden paths: assessing the
contributions of syntax, semantics, and plausibility in resolving ambiguities,
Language and Cognitive Processes, 11, 283334.
Obler, L. K. and Gjerlow, K. (1999), Language and the Brain, Cambridge: Cambridge
University Press.
Ogden, J. A. (1988), Language and memory functions after long recovery periods in
left hemispherectomised subjects, Neuropsychologia, 26, 64559.
Osterhout, L. and Nicol, J. (1999), On the distinctiveness, independence and time
course of the brain responses to syntactic and semantic anomalies, Language and
Cognitive Processes, 14, 283317.
Paivio, A. (1986), Mental Representations: A Dual Coding Approach, New York:
Oxford University Press.
Bibliography 251
Paradis, M. (1977), Bilingualism and aphasia, in H. Whitaker (ed.), Studies in

Neurolinguistics (Vol. 3), New York: Academic Press, pp. 65121.
Paradis, M. (1998), Acquired aphasia in bilingual speakers, in M. T. Sarno (ed.),
Acquired Aphasia (3rd edn), San Diego: Academic Press, pp. 53149.
Paradis, M., Goldblum, M. C. and Abidi, R. (1982), Alternate antagonism with par-
adoxical translation in two bilingual aphasic patients, Brain and Language, 15,
5569.
Park, T. Z. (1978), Plurals in child speech, Journal of Child Language, 5, 23750.
Paterson, S. J., Brown, J. H., Gsdl, M. H., Johnson, M. H. and Karmilo-Smith, A.
(1999), Cognitive modularity and genetic disorders, Science, 286, 23558.
Pawley, A. and Syder, F. H. (1983), Two puzzles for linguistic theory: nativelike selec-
tion and nativelike uency, in J. C. Richards and R. W. Schmidt (eds), Language and
Communication, London: Longman, pp. 191226.
Prez-Pereira, M. (1994), Imitations, repetitions, routines and the childs analysis of
language: insights from the blind, Journal of Child Language, 21, 31737.
Prez-Pereira, M. and Castro, J. (1997), Language acquisition and the compensation
of visual decit: new comparative data on a controversial topic, British Journal of
Developmental Psychology, 15, 43959.
Prez-Pereira, M. and Conti-Ramsden, G. (1999), Language Development and Social
Interaction in Blind Children, Hove: Psychology Press.
Peters, A. M. (1977), Language learning strategies: does the whole equal the sum of
the parts?, Language, 53, 56073.
Peters, A. M. (1983), The Units of Language Acquisition, Cambridge: Cambridge
University Press.
Peters, A. M. (1985), Language segmentation: operating principles for the perception
and analysis of language, in D. I. Slobin (ed.), The Cross-Linguistic Study of
Language Accquisition. Vol. 2: Theoretical Issues, Hillsdale, NJ: Lawrence Erlbaum,
pp. 102967.
Peters, A. M. (1995), Strategies in the acquisition of syntax, in P. Fletcher and B.
MacWhinney (eds), The Handbook of Child Language, Oxford: Blackwell, pp.
46282.
Peters, A. M. (2001), Filler syllables: what is their status in emerging grammar?,
Peters, A. M. and Menn, L. (1993), False starts and ller syllables: ways to learn gram-
matical morphemes, Language, 69, 74277.
Peterson, R. and Burgess, C. (1993), Syntactic and semantic processing during idiom
comprehension: neurolinguistic and psycholinguistic dissociations, in C. Cacciari
and P. Tabossi (eds), Idioms: Processing, Structure and Interpretation, Hillsdale, NJ:
Petitto, L. A. (1987), On the autonomy of language and gesture: evidence from the
acquisition of personal pronouns in American Sign Language, Cognition, 27, 152.
Piattelli-Palmarini, M. (1989), Evolution, selection and cognition: from learning to
parameter setting in biology and the study of language, Cognition, 31, 144.
Pick, A. (1973) [1913], Aphasia, trans. and ed. J. Brown, Springeld, IL: Charles C.
Thomas.
Pine, J. (1994), The language of primary caregivers, in C. Gallaway and B. J. Richards
(eds), Input and Interaction in Language Acquisition, Cambridge: Cambridge
University Press, pp. 1537.
Pine, J. M. and Lieven, E. V. (1993), Reanalysing rote-learned phrases: individual

dierences in the transition to multi-word speech, Journal of Child Language, 20,
55171.
Pinker, S. (1984), Language Learnability and Language Development, Cambridge, MA:
Pinker, S. (1987), The bootstrapping problem in language acquisition, in B.
MacWhinney (ed.), Mechanisms of Language Acquisition, Hillsdale, NJ: Lawrence
Erlbaum, pp. 399441.
Pinker, S. (1991), Rules of language, Science, 253, 5305.
Pinker, S. (1995), Facts about human language relevant to its evolution, in J.-P.
Changeux and J. Chavaillon (eds), Origins of the Human Brain, Oxford: Clarendon
Press, pp. 26283.
Pinker, S. (1998), Words and rules, Lingua, 106, 21942.
Pinker, S. (1999), Words and Rules: The Ingredients of Language, London: Weidenfeld
and Nicolson.
Pinker, S. and Bloom, P. (1990), Natural language and natural selection, Behavioral
Pinker, S. and Prince, A. (1988), On language and connectionism: analysis of a par-
allel distributed processing model of language acquisition, Cognition, 28, 73193.
Pinker, S. and Prince, A. (1992), Regular and irregular morphology and the psycho-
logical status of rules of grammar, Berkeley Linguistics Society, 17, 23051.
Pizzuto, E. and Caselli, M. C. (1992), The acquisition of Italian morphology: impli-
cations for models of language development, Journal of Child Language, 19,
491557.
Plunkett, K. and Marchman, V. (1991), U-shaped learning and frequency eects in a
multi-layered perceptron: implications for child langauge acquisition, Cognition, 38,
43102.
Plunkett, K. and Marchman, V. (1993), From rote learning to system building: acquir-
ing verb morphology in children and connectionist nets, Cognition, 48, 2169.
Plunkett, K. and Nakisa, R. C. (1997), A connectionist model of the Arabic plural
system, Language and Cognitive Processes, 12, 80736.
Pollack, I. and Pickett, J. M. (1964), Intelligibility of excerpts from uent speech: audi-
tory vs. structural context, Journal of Verbal Learning and Verbal Behavior, 3, 7984.
Posner, M. I., Petersen, S., Fox, P. and Raichle, M. (1988), Localization of cognitive
operations in the human brain, Science, 240, 162731.
Prasada, S. and Pinker, S. (1993), Generalization of regular and irregular morpholog-
ical patterns, Language and Cognitive Processes, 8, 156.
Premack, D. (1980), Representational capacity and the accessibility of knowledge: the
case of chimpanzees, in M. Piattelli-Palmarini (ed.), Language and Learning: The
Debate Between Jean Piaget and Noam Chomsky, London: Routledge and Kegan
Paul, pp. 20321.
Preuss, T. M. (2000), Whats human about the human brain?, in M. S. Gazzaniga
(ed.), The New Cognitive Neurosciences, Cambridge, MA: MIT Press, pp. 121934.
Pullum, G. K. and Scholz, B. C. (2002), Empirical assessment of stimulus poverty
arguments, Linguistic Review, 19, 950.
Pulvermller, F. (1998), On the matter of rules: past-tense formation and its signi-
cance for cognitive neuroscience, Network, 9, R1R52.
Putnam, H. (1988), Representation and Reality, Cambridge, MA: MIT Press.
Bibliography 253
Quirk, R., Greenbaum, S., Leech, G. and Svartnik, J. (1985), A Comprehensive

Grammar of the English Language, London: Longman.
Radford, A. (1990), Syntactic Theory and the Acquisition of English Syntax: The
Nature of Early Child Grammars of English, Oxford: Blackwell.
Ramscar, M. (2002), The role of meaning in inection: why the past tense does not
require a rule, Cognitive Psychology, 45, 4594.
Rayner, K., Carlson, M. and Frazier, L. (1983), The interaction of syntax and seman-
tics during sentence processing, Journal of Verbal Learning and Verbal Behavior, 22,
35874.
Redington, M. and Chater, N. (1998), Connectionist and statistical approaches to lan-
guage acquisition: a distributional perspective, Language and Cognitive Processes,
13, 12992.
Regier, T. (1996), The Human Semantic Potential: Spatial Language and Constrained
Connectionism, Cambridge, MA: MIT Press.
Richards, B. J. (1990), Language Development and Individual Dierences: A Study of
Auxiliary Verb Learning, Cambridge: Cambridge University Press.
Richards, M. (1979), Adjective ordering in the language of young children: an experi-
mental investigation, Journal of Child Language, 6, 25378.
Rosch, E. (1975), Cognitive reference points, Cognitive Psychology, 7, 53247.
Rosch, E. (1977), Human categorization, in N. Warren (ed.), Studies in Cross-Cultural
Psychology (Vol. 1), London: Academic Press, pp. 149.
Rowland, C. F. and Pine, J. M. (2000), Subject-auxiliary inversion errors and wh-
question acquisition: what children do know?, Journal of Child Language, 27,
15781.
Rubino, R. B. and Pine, J. M. (1998), Subjectverb agreement in Brazilian Portuguese:
what low error rates hide, Journal of Child Language, 25, 3559.
Rudzka-Ostyn, B. (1985), Metaphoric processes in word formation: the case of pre-
xed verbs, in W. Paprott and R. Dirven (eds), The Ubiquity of Metaphor.
Metaphor in Language and Thought, Amsterdam: John Benjamins, pp. 20941.
Rudzka-Ostyn, B. (ed.) (1988), Topics in Cognitive Linguistics, Amsterdam: John
Benjamins.
Rumelhart, D. E. and McClelland, J. L. (1986), On learning the past tenses of English
verbs, in D. E. Rumelhart and J. L. McClelland (eds), Parallel Distributed
Processing: Explorations in the Microstructure of Cognition. Vol. 2: Psychological
and Biological Models, Cambridge, MA: MIT Press, pp. 21671.
Sachs, J. (1983), Talking about the there and then: the emergence of displaced refer-
ence in parentchild discourse, in K. E. Nelson (ed.), Childrens Language (Vol. 4),
Sachs, J., Bard, B. and Johnson, M. L. (1981), Language learning with restricted input:
case studies of two hearing children of deaf parents, Applied Psycholinguistics, 2,
3354.
Samuel, A. G. (1986), The role of the lexicon in speech perception, in E. B. Schwab
and H. C. Nusbaum (eds), Pattern Recognition by Humans and Machines. Vol. 1:
Speech Perception, Orlando: Academic Press, pp. 89111.
Savage, C., Lieven, E., Theakston, A. and Tomasello, M. (submitted), Testing the
abstractness of childrens linguistic representations: lexical and structural priming of
syntactic constructions in young children.
Schneider, W. and Bjorklund, D. F. (1998), Memory, in D. Kuhn and R. S. Siegler
(eds), The Handbook of Child Psychology. Vol. 2: Cognition, Perception and

Language, New York: John Wiley and Sons, pp. 467522.
Schreuder, R., de Jong, N., Krott, A. and Baayen, H. (1999), Rules and rote: beyond
the linguistic either-or fallacy, Behavioral and Brain Sciences, 22, 10389.
Seidenberg, M. S. and Hoener, J. H. (1998), Evaluating behavioral and neuroimag-
ing data on past tense processing, Language, 74, 10422.
Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M. and Bienkowski, M. (1982),
Automatic access of the meanings of ambiguous words in context: some limitations
of knowledge-based processing, Cognitive Psychology, 14, 489537.
Senghas, A. and Coppola, M. (2001), Children creating language: how Nicaraguan
Sign Language acquired a spatial grammar, Psychological Science, 12, 3238.
Sereno, J., Zwitserlood, P. and Jongman, A. (1999), Entries and operations: the great
divide and the pitfalls of form frequency, Behavioral and Brain Sciences, 22, 1039.
Sereno, J. A. and Jongman, A. (1997), Processing of English inectional morphology,
Memory and Cognition, 25, 42537.
Silverberg, R. and Gordon, H. W. (1979), Dierential aphasia in two bilingual indi-
viduals, Neurology, 29, 515.
Simpson, G. B. (1994), Context and the processing of ambiguous words, in M. A.
Gernsbacher (ed.), Handbook of Psycholinguistics, San Diego: Academic Press, pp.
35974.
Sinclair, J. (1987), Introduction, in J. Sinclair (ed.), Collins COBUILD English
Language Dictionary, London: HarperCollins, pp. xvxxi.
Sinha, C. and Kuteva, T. (1995), Distributed spatial semantics, Nordic Journal of
Sinha, C., Thorseng, L. A., Hayashi, M. and Plunkett, K. (1994), Comparative spatial
semantics and language acquisition: evidence from Danish, English, and Japanese,
Journal of Semantics, 11, 25387.
Slobin, D. I. (1973), Cognitive prerequisites for the development of grammar, in C.
A. Ferguson and D. I. Slobin (eds), Studies in Child Language Development, New
York: Holt, Rinehart and Winston, pp. 175208.
Slobin, D. I. (1991), Aphasia in Turkish: speech production in Brocas and Wernickes
patients, Brain and Language, 41, 14964.
Smith, L. B., Jones, S. S. and Landau, B. (1992), Count nouns, adjectives and percep-
tual properties in childrens novel word interpretation, Developmental Psychology,
28, 27386.
Smith, N. and Tsimpli, I. M. (1991), Linguistic modularity? A case study of a savant
linguist, Lingua, 84, 31551.
Smith, N. and Tsimpli, I. M. (1995), The Mind of a Savant: Language Learning and
Modularity, Oxford: Blackwell.
Smoczynska, M. (1985), The acquisition of Polish, in D. I. Slobin (ed.), The
Crosslinguistic Study of Language Acquisition. Vol. 1: The Data, Hillsdale, NJ:
Smoczynska, M. (1998), Krakowskie dane jezykowe dzieci w systemie CHILDES, in
M. Smoczynska (ed.), Studia z psychologii rozwojowej i psycholingwistyki, Krakw:
Universitas, pp. 28396.
Snow, C. (1986), Conversations with children, in P. Fletcher and M. Garman (eds),
Language Acquisition (2nd edn), Cambridge: Cambridge University Press, pp.
6989.
Bibliography 255
Snow, C. E. (1981), The uses of imitation, Journal of Child Language, 8, 20512.

Snow, C. E. (1983), Saying it again: the role of expanded and deferred imitations in
language acquisition, in K. E. Nelson (ed.), Childrens Language (Vol. 4), Hillsdale,
NJ: Lawrence Erlbaum, pp. 2958.
Snow, C. E. (1995), Issues in the study of input: netuning, universality, individual and
developmental dierences, and necessary causes, in P. Fletcher and B. MacWhinney
(eds), The Handbook of Child Language, Oxford: Blackwell, pp. 18093.
Sonnenstuhl, I., Eisenbeiss, S. and Clahsen, H. (1999), Morphological priming in the
German lexicon, Cognition, 72, 20336.
Spivey, M. J., Tanenhaus, M. K., Eberhard, K. M. and Sedivy, J. C. (2002), Eye move-
ments and spoken language comprehension: eects of visual context on syntactic
ambiguity resolution, Cognitive Psychology, 45, 44781.
Springer, S. (1979), Speech perception and the biology of language, in M. Gazzaniga
(ed.), Handbook of Behavioral Neurobiology. Vol. 2: Neuropsychology, New York:
Plenum Press, pp. 15377.
Stemberger, J. P. and MacWhinney, B. (1986), Frequency and the lexical storage of reg-
ularly inected forms, Memory and Cognition, 14, 1726.
Stemberger, J. P. and MacWhinney, B. (1988), Are lexical forms stored in the lexicon?,
in M. Hammond and M. Noonan (eds), Theoretical Morphology: Approaches in
Modern Linguistics, London: Academic Press, pp. 10116.
Stojanovik, V., Perkins, M. and Howard, S. (submitted), Williams syndrome and spe-
cic language impairment do not support claims for developmental double dissoci-
ations.
Stromswold, K. (1995), The acquisition of subject and object wh-questions,
Language Acquisition, 4, 548.
Stromswold, K. (1996), Genes, specicity, and the lexical/functional distinction in lan-
guage acquisition, Behavioral and Brain Sciences, 19, 6489.
Stromswold, K. (2000), The cognitive neuroscience of language acquisition, in M. S.
Gazzaniga (ed.), The New Cognitive Neurosciences, Cambridge, MA: MIT Press, pp.
90932.
Stromswold, K. (2001), The heritability of language: a review and metaanalysis of
twin, adoption and linkage studies, Language, 77, 647723.
Sutton-Spence, R. and Woll, B. (1999), The Linguistics of British Sign Language: An
Introduction, Cambridge: Cambridge University Press.
Svorou, S. (1994), The Grammar of Space, Amsterdam: John Benjamins.
Tager-Flusberg, H. (1986), Constraints on the representation of word meaning: evi-
dence from autistic and mentally retarded children, in S. A. Kuczaj and M. Barrett
(eds), The Development of Word Meaning, New York: Springer-Verlag, pp. 6981.
Talmy, L. (1983), How language structures space, in H. Pick and L. Acredolo (eds),
Spatial Orientation: Theory, Research and Application, New York: Plenum Press, pp.
22582.
Talmy, L. (1985), Force dynamics in language and thought, Parasession on Causatives
and Agentivity, 21st Regional Meeting, Chicago Linguistic Society, 293337.
Tanenhaus, M. K., Leiman, J. M. and Seidenberg, M. S. (1979), Evidence for multiple
stages in the processing of ambiguous words in syntactic contexts, Journal of Verbal
Learning and Verbal Behavior, 18, 42740.
Taylor, J. R. (1989), Linguistic Categorization: Prototypes in Linguistic Theory, Oxford:
Clarendon Press.
Taylor, J. R. (2002), Cognitive Grammar, Oxford: Oxford University Press.

Tesak, J. (1994), Cognitive load and the processing of grammatical items,
Neurolinguistics, 8, 438.
Thal, D., Bates, E. and Bellugi, U. (1989), Language and cognition in two children
with Williams syndrome, Journal of Speech and Hearing Research, 32, 489500.
Thal, D. J., Bates, E., Zappia, M. J. and Oroz, M. (1996), Ties between lexical and
grammatical development: evidence from early talkers, Journal of Child Language,
23, 34968.
Theakston, A., Lieven, E., Pine, J. and Rowland, C. (2000), The role of performance
limitations in the acquisition of mixed verb-argument structure at stage 1, in M.
Perkins and S. Howard (eds), New Direction in Language Development and Disorders,
New York: Plenum Press, pp. 11928.
Thomas, M. S. C., Grant, J., Barham, Z., Gsdl, M., Laing, E., Lakusta, L., Tyler, K.,
Grice, S., Paterson, S. and Karmilo-Smith, A. (2001), Past tense formation in
Williams syndrome, Language and Cognitive Processes, 16, 14376.
Thornton, R. and Crain, S. (1994), Successful cyclic movement, in T. Hoekstra and
B. D. Schwartz (eds), Language Acquisition Studies in Generative Grammar,
Amsterdam: John Benjamins, pp. 21552.
Todd, P. (1982), Tagging after red herrings: evidence against the processing capacity
explanation in child language, Journal of Child Language, 9, 99114.
Todd, P. and Aitchison, J. (1980), Learning language the hard way, First Language, 1,
12240.
Tomasello, M. (1987), Learning to use prepositions: a case study, Journal of Child
Language, 14, 7998.
Tomasello, M. (1992), First Verbs: A Case Study of Early Grammatical Development,
Cambridge: Cambridge University Press.
Tomasello, M. (1995), Language is not an instinct, Cognitive Development, 10,
13156.
Tomasello, M. (1999), The Cultural Origins of Human Cognition, Cambridge, MA:
Tomasello, M. (2000), Do young children have adult syntactic competence?,
Tomasello, M. (2003), Constructing a Language: A Usage-Based Theory of Child
Language Acquisition, Cambridge, MA: Harvard University Press.
Tomasello, M. and Abbot-Smith, K. (2002), A tale of two theories: response to
Fisher, Cognition, 83, 20714.
Tomasello, M. and Todd, J. (1983), Joint attention and lexical acquisition style, First
Tomasello, M., Akhtar, N., Dodson, K. and Rekau, L. (1997), Dierential productiv-
ity in young childrens use of nouns and verbs, Journal of Child Language, 24,
37387.
Tomblin, J. B. and Buckwalter, P. R. (1998), Heritability of poor language achievement
among twins, Journal of Speech Language and Hearing Research, 41, 18899.
Traugott, E. C. (1978), On the expression of spatio-temporal relations in language, in
J. Greenberg, C. Ferguson and E. Moravcsik (eds), Universals of Human Language
(Vol. 3), Stanford, CA: Stanford University Press, pp. 369400.
Trueswell, J. C. (1996), The role of lexical frequency in syntactic ambiguity resolution,
Journal of Memory and Language, 35, 56685.
Bibliography 257
Trueswell, J. C., Tanenhaus, M. K. and Garnsey, S. M. (1994), Semantic inuences on

parsing: use of thematic role information in syntactic ambiguity resolution, Journal
of Memory and Language, 33, 285318.
Trueswell, J. C., Tanenhaus, M. K. and Kello, C. (1993), Verb-specic constraints in
sentence-processing: separating eects of lexical preference from garden-paths,
Journal of Experimental Psychology: Learning Memory and Cognition, 19, 52853.
Ulatowska, H. K., Sadowska, M. and Kadzielowa, D. (2001), A longitudinal study of
agrammatism in Polish: a case study, Journal of Neurolinguistics, 14, 32136.
Ullman, M., Bergida, R. and OCraven, K. M. (1997), Distinct fMRI activation pat-
terns for regular and irregular past tense, NeuroImage, 5, S549.
Ullman, M. T. (1999), Acceptability ratings of regular and irregular past tense forms:
evidence for a dual-system model of language from word frequency and phonolog-
ical neighborhood eects, Language and Cognitive Processes, 14, 4767.
Ullman, M. T., Corkin, S., Coppola, M., Hickok, G., Growdon, J. H., Koroshetz, W.
J. and Pinker, S. (1997), A neural dissociation within language: evidence that the
mental dictionary is part of declarative memory, and that grammatical rules are pro-
cessed by the procedural system, Journal of Cognitive Neuroscience, 9, 227.
Urwin, C. (1984), Communication in infancy and the emergence of language in blind
children, in R. L. Schiefelbusch and J. Pickar (eds), The Acquisition of Communicative
Competence, Baltimore: University Park Press, pp. 479524.
van Hoek, K. (1997), Anaphora and Conceptual Structure, Chicago: Chicago
University Press.
van Hout, A. (1991), Outcome of acquired aphasia in childhood: prognosis factors,
in I. Pavo Martins, A. Castro-Caldas, H. R. van Dongen and A. van Hout (eds),
Acquired Aphasia in Children: Acquisition and Breakdown of Language in the
Developing Brain, Dordrecht: Kluwer, pp. 1639.
van Mourik, M., Verschaeve, M., Boon, P., Paquier, P. and Vanharskamp, F. (1992),
Cognition in global aphasia: indicators for therapy, Aphasiology, 6, 4919.
Van Petten, C., Coulson, S., Rubin, S., Plante, E. and Parks, M. (1999), Time course
of word identication and semantic integration in spoken language, Journal of
Experimental Psychology: Learning Memory and Cognition, 25, 394417.
Vargha-Khadem, F. and Polkey, C. E. (1992), A review of cognitive outcome after
hemidecortication in humans, in F. D. Rose and D. A. Johnson (eds), Recovery from
Brain Damage: Reections and Directions, New York: Plenum Press, pp. 13751.
Vargha-Khadem, F., Watkins, K., Alcock, K., Fletcher, P. and Passingham, R. (1995),
Praxic and nonverbal cognitive decits in a large family with a genetically transmit-
ted speech and langauge disorder, Proceedings of the National Academy of Sciences
of the United States of America, 92, 9303.
Verhagen, A. (forthcoming), Constructions of Intersubjectivity, Oxford: Oxford
University Press.
Volterra, V., Caparci, O., Pezzini, G., Sabbadini, L. and Vicari, S. (1996), Linguistic
abilities in Italian children with Williams syndrome, Cortex, 32, 66377.
Warburton, E., Wise, R. J. S., Price, C. J., Weiller, C., Hadar, U., Ramsay, S. and
Frackowiak, R. S. J. (1996), Noun and verb retrieval by normal subjects: studies
with PET, Brain, 119, 15979.
Wason, P. C. and Reich, S. S. (1979), A verbal illusion, Quarterly Journal of
Experimental Psychology, 31, 5917.
Wayland, S. C., Berndt, R. S. and Sandson, J. R. (1996), Aphasic patients sensitivity
to structural and meaning violations when monitoring for nouns and verbs in sen-
tences, Neuropsychology, 10, 50416.
Wegener, H. (1994), Variation in the acquisition of German plural morphology by
second language learners, in R. Tracy and E. Lattey (eds), How Tolerant is Universal
Grammar? Essays on Language Variability and Language Variation, Tbingen:
Niemeyer, pp. 26794.
Wegener, H. (1999), Die Pluralbildung im Deutschen: ein Versuch im Rahmen der
Optimalittstheorie, Linguistik online, 4, http://viadrina.euv-frankfurt-o.de/~
wjournal/3{I99/wegener.html.
Wells, C. G. (1979), Learning and using the auxiliary verb in English, in V. Lee (ed.),
Language Development, London: Croom Helm, pp. 25070.
Wells, G. (1985), Language Development in the Preschool Years, Cambridge:
Westfal, S. (1956), A Study in Polish Morphology: The Genitive Singular Masculine, The
Hague: Mouton.
Whiten, A., Custance, D. M., Gomez, J. C., Teixidor, P. and Bard, K. A. (1996),
Imitative learning of articial fruit processing in children (Homo sapiens) and chim-
panzees (Pan troglodytes), Journal of Comparative Psychology, 110, 314.
Wilson, B. and Peters, A. M. (1988), What are you cookin on a hot? A three-year-old
blind childs violation of universal constraints on constituent movement, Language,
64, 24973.
Wise, R. J. S., Chollet, F., Hadar, U., Friston, K., Honer, E. and Frackowiak, R.
(1991), Distribution of cortical networks in word comprehension and word
retrieval, Brain, 114, 180317.
Worden, R. (1998), The evolution of language from social intelligence, in J. R.
Hurford, M. Studdert-Kennedy and C. Knight (eds), Approaches to the Evolution of
Language: Social and Cognitive Bases, Cambridge: Cambridge University Press, pp.
14866.
Wulfeck, B. and Bates, E. (1991), Dierential sensitivity to errors of agreement and
word order in Brocas aphasia, Journal of Cognitive Neuroscience, 3, 25872.
Wulfeck, B., Bates, E. and Capasso, R. (1991), A crosslinguistic study of grammatical-
ity judgments in Brocas aphasia, Brain and Language, 41, 31136.
Yamada, J. E. (1990), Laura: A Case for Modularity of Language, Cambridge, MA:
MIT Press.
Zattore, R., Evans, A. C., Meyer, E. and Gjedde, A. (1992), Lateralization of phonetic
and pitch discrimination in speech processing, Science, 256, 8469.

Language Mind and Brain Dabrowska Libros

Uploaded by

Copyright:

Available Formats

Language Mind and Brain Dabrowska Libros

Uploaded by

Document Information

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

Language Mind and Brain Dabrowska Libros

Uploaded by

Copyright:

Available Formats

What is the book about based on the introduction and table of contents?

What is the book about based on the introduction and table of contents?

What are some of the topics discussed in Part I of the book?

What are some of the topics discussed in Part I of the book?

DABROWSKA pre Make up 26/3/04 4:32 pm Page i

Language, Mind and Brain

To the memory of Mieczysl-aw Dabrowski

Language, Mind and

EDINBURGH UNIVERSITY PRESS

Ewa Dabrowska, 2004

Edinburgh University Press Ltd

Typeset in 10/12 Times New Roman

ISBN 0 7486 1474 5 (hardback)

The right of Ewa Dabrowska

List of figures viii

PART I THE BASIC SPECIFICATIONS 5

Chapter 2 Language processing: speed and flexibility 7

Chapter 3 Language acquisition: robustness 29

Chapter 4 Language in the brain 40

Chapter 5 Language and other cognitive processes 50

vi Language, mind and brain

Chapter 6 Biological underpinnings 58

PART II THE BUILDING BLOCKS OF LANGUAGE 79

Chapter 8 On rules and regularity 116

8. Interim conclusions 147

Chapter 9 Syntactic Constructions 159

Chapter 10 The cognitive enterprise 203

7.1 Is the banana in the bowl? 85

8.1 Circumstances in which memory is not accessed and the default

I would like to thank my teachers Roman Kalisz and Tomasz Krzeszowski,

Linguistics is often dened as the branch of cognitive science dealing with

2 Language, mind and brain

organisation, so each of these areas is explored from a developmental perspec-

information about systematic exceptions to generalisations proposed by lin-

2 Language processing: speed and

1. The complexity of the task

Most people including many language professionals take language for

(1) Tess blew her nose on a paper tissue.

8 Language, mind and brain

1. moved (of air particles), as in The wind blew in my face;

13. having a particular colour;

Language processing: speed and flexibility 9

1. a part of the body;

It can also have a range of verbal meanings:

6. (of an aircraft) to come down steeply;

The word paper also has several dierent senses:

In addition to these multiply ambiguous open-class words, our sentence

(2) a. All the books on politics are on this shelf.

10 Language, mind and brain

f. The house is on the river.

1. to refer to a specic member of the class designated by the following noun;

Language processing: speed and flexibility 11

12 Language, mind and brain

Language processing: speed and flexibility 13

Casual speech is also noisy in a metaphorical sense, in that it contains a

In spite of this complexity, humans process language extremely fast. In infor-

14 Language, mind and brain

Language processing: speed and flexibility 15

16 Language, mind and brain

1. The left-anterior negativity (LAN) is a negative-amplitude wave in the left