Lexical Semantics

2
Lexical semantics
James Pustejovsky
2.1 Introduction 33
2.2 The history of lexical semantics 36
2.3 Issues in lexical semantics 39
2.4 Event semantics 48
2.5 Lexical decomposition 52
2.6 Semantic roles 55
2.7 Qualia structure 59
2.8 Type theory and the lexicon 60
2.9 Open issues in lexical semantics 63
2.1 Introduction
Lexical semantics is the study of what words mean and how their meanings
contribute to the compositional interpretation of natural language utter-
ances. The lexicon can be seen as that component of the grammar that
encodes both the information required for composition in the syntax and
the knowledge for multiple levels and types of semantic interpretation. Lex-
ical entries are richly structured objects that act as triggers both to com-
positional operations and to entailments and implicatures in the context
of larger discourses. Because any semantic interpretation requires access to
knowledge about words, the lexicon of a grammar must provide a system-
atic and efficient way of encoding the information associated with words in
a language.
Four key questions arise when determining how to model the meanings
conveyed by words:
(i) What are the semantic components that constitute word meaning?
(ii) How are word meanings differentiated and how are they related to each
other?
Downloaded from https://www.cambridge.org/core. Access paid by the UCSB Libraries, on 16 Jun 2018 at 15:36:16, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/CBO9781139236157.003
34 JA M E S PU ST E J OV S KY
(iii) How does the meaning of individual words drive the compositional pro-
cess to make semantically coherent sentences?
(iv) When is a component of word meaning considered “lexical” rather
than “world” knowledge?
As the linguistic phenomena associated with lexical semantics become

better understood, several theoretical assumptions have emerged across
most current models of word meaning. These can be summarized roughly
as follows:
r Lexical meaning involves a kind of componential analysis, either through
predicative primitives or a system of types.
r The selectional properties of words must be explained in terms of the
lexical semantics of predication.
r An understanding of the semantics of nouns and the contribution that
verbal arguments play in composition is crucial for an adequate model of
how lexical semantics contributes to compositionality.
As we will see below, the first point makes an explicit connection between
predicate decomposition theories (such as Lakoff, 1965/1970; Levin and
Rappaport-Hovav, 1995) and type-theoretic approaches to lexical semantics
(Dowty, 1979; Pustejovsky, 1995; Davis and Koenig, 2000; Asher and Puste-
jovsky, 2006; Asher, 2011). This in turn directly influences the manner in
which selectional constraints are encoded. Finally, we will observe the cen-
tral role of Aktionsarten and event typing in the determination of sentence
meaning in composition.
There are essentially four strategies to lexical specification that have been
adopted in the literature. These approaches can be defined in terms of
how, or whether, they provide an intermediate level of interpretation from
expressions in the object language to their denotations in the model. The
interface defines the logical forms associated with lexical items in the lan-
guage, and it is these expressions which are interpreted in the model.
The four approaches can be defined as follows:
Atomic predication: No interface language is provided, and there is direct

interpretation of an object expression into the model.
Relational models: The interface is a relational structure, which is inter-
preted in the model.
Feature-based decomposition: Component-based features are used to clas-
sify an expression in the object language into distinct concepts in the
model.
Structural decomposition: Component-based features are organized as a
graph structure, with associated compositional interpretations in the
model.
Atomic predication refers to non-decompositional approaches to lexi-

cal semantic interpretation, such as the early type-theoretic semantics of
Lexical semantics 35
Montague (1974), where the lexical items are primitives in the object lan-
guage, or the “language of thought” proposed by Fodor (1975).
This view assumes a very tight coupling between the lexical semantics of
a word and the syntactic projection associated with it. One consequence of
this position is that there are as many lexical entries for a word as there are
senses for it in the object language.
Relational models also start with the assumption that words are treated
as primitives, but unlike atomic predication theory, they can have arbitrar-
ily complex relational structures that facilitate semantic inferences. Such
approaches are not as strict about introducing arguments that are visibly
expressed in the syntax of the language, as is done with atomic predica-
tion models. Perhaps the best-known example of this strategy is Davidson’s
addition of the event variable to action predicates (Davidson, 1967a), as well
as most subsequent work assuming an event variable for eventuality predi-
cates in language. Computational models for linguistic inference also often
invoke this strategy in order to perform reasoning within established logi-
cal systems. Hobbs et al. (1993), for example, working within a framework of
first-order abductive inference, adds any additional parameters to the argu-
ment structure of a predicate that are needed for an inference.
Feature-based decomposition has been used for discriminative analysis
of natural language semantics since Nida (1949), and then, significantly
in the 1960s, when the Katz and Fodor (1963) and Katz and Postal (1964)
models within early generative grammar gained currency within the field.
All expressions in the object language are decomposed into sets of binary-
valued features, distinguishing concepts such as gender, number, age, mar-
ital status, and so on. Recently, some of these ideas have come back into
vogue, where vector-based representations of word meaning have emerged
as a way to handle some long-standing problems in the computational
interpretation of language. These distributional semantic models utilize
far more sophisticated techniques for identifying word distributions and
computing similarity and relatedness with them (Schütze, 1993; Schütze,
1995; Landauer and Dumais, 1997; Padó and Lapata, 2007; Erk and Padó,
2008).
Finally, structural decomposition is the approach currently adopted by
many lexical semanticists working in the interface between syntax and
semantics. Words are defined as restricted algebraic structures, with primi-
tive predicates as atomic elements. This approach has been adopted broadly,
from Dowty (1979) in his model-theoretic interpretation of generative
semantics (Lakoff, 1965/1970), to Van Valin Jr. (2005), Jackendoff’s Concep-
tual Structure (Jackendoff, 1983, 1990), and variants of this model (Levin
and Rappaport-Hovav, 1995, 2005). Both Generative Lexicon Theory (Puste-
jovsky and Boguraev, 1993; Pustejovsky, 1995) and semantic interpretations
for Head-Driven Phrase Structure Grammar (HPSG) (Davis and Koenig, 2000;
Ginzburg and Sag, 2000) can also be seen as assuming a rich, structural
decomposition as the foundation for their models.
In this entry, after a brief history of research in lexical semantics, we

review the major issues confronting models of word meaning. Then we
examine more closely the meaning components of individual words, begin-
ning with word classes and types, and moving on to different strategies for
encoding lexical meaning, focusing on polysemy and argument selection.
Next we consider the role played by event semantics in the determination
of meaning and survey approaches to decomposition and semantic roles.
After a brief introduction to qualia structure, we conclude with an exam-
ination of the ways that word meanings contribute to the compositional
mechanisms responsible for building sentence meanings.
2.2 The history of lexical semantics
Interest in the meaning of words goes back as far as philosophical thought

and speculation about the natural world. Aristotle, for example, in his De
Interpretatione (McKeon, 1968) outlines a broad theory of language mean-
ing, framed within a context of determining truth in the service of reason-
ing. As part of his general characterization of nature, Aristotle proposes a
classification of thoughts and the words that convey them using a system-
atic set of dimensions called aitia. Applied to verbs and their correspond-
ing activities, this allows one to distinguish between telic (goal-completive)
and atelic (non-goal completive) eventualities. This has both influenced work
in lexical aspect and Aktionsarten within linguistics and provided much
of the conceptual background for qualia structure (Pustejovsky, 1995), see
Section 2.7.
With the exception of Locke (1690), Hume (1938), and Reid (1764), there
was very little in the way of serious or systematic theorizing about word
meaning until late in the nineteenth century. At this time, language
researchers in Germany and France began focusing on lexical semantics
from a psychological perspective, looking at the relation between words and
concepts. Bréal (1897), for example, considered polysemy (a term he coined)
to be a necessary creative component of language, claiming that this phe-
nomenon better than most others in semantics illustrates the cognitive and
conceptualizing force of the human species. Similarly, German semasiolo-
gists viewed the lexicon, word meaning, and mechanisms of polysemy as
illustrative of the “life force” of human language (see Erdmann, 1900).
In the early twentieth century, European Structuralists introduced the
distinction between syntagmatic and paradigmatic processes. Syntagmatic
relations refer to the influence of “horizontal” elements on a word or phrase
(co-occurrence), in contradistinction to paradigmatic processes, which refer
to “vertical” or alternative substitutions in a phrase. The terms had sig-
nificant currency in early and mid-twentieth-century linguistics from de
Saussure on and helped to define the formal study of syntax as widely
practiced today. The Structuralists saw paradigmatic relations encoded in
the various lexical systems of a language, and this was elaborated into a
framework of componential analysis for language meaning (Nida, 1949;
Jakobson, 1973). Componential analysis, used by anthropologists to study
kinship terminology, is a method for breaking down the meanings of lexical
items into a set of features, thereby illustrating similarities and differences
of meaning between the items. The goal of such analysis was simply to
classify the lexical items in the language with some finite set of features, its
ultimate contrastive elements. These contrastive elements are then struc-
tured in a matrix, allowing for dimensional analyses and generalizations
to be made about lexical sets occupying the cells in the matrix. Hjelmslev
(1943/1961), for example, decomposed lexical items into “paradigms” which
pattern the same way distributionally in the language. The componential
analysis of lexical items entails a decomposition into distinctive features:
man as [+adult, +male], woman as [+adult, −male], girl as [−adult, −male],
and so forth. Similar distinctions are made in Nida (1949) for distinguishing
morphological patterns.
In the early days of generative linguistics, many of the ideas of the Struc-
turalists found their way into the first formulations of lexical knowledge for
transformational grammars. Of particular importance was Katz and Fodor’s
(1963) and Katz and Postal’s (1964) theory of feature-based semantics, where
the meanings of words were composed entirely of sets of features with
Boolean values. In line with Chomsky (1957), the role of meaning was lim-
ited: to detect semantic anomalies and determine the number of readings
associated with a sentence. The theory offered proposals for the decomposi-
tion of sentences and lexical items, with explicit rules for linking items to
syntactic structure. While influential in the short term, this theory had no
adequate theory of compositionality and was seen to be too weak as a model
for natural language semantics (see Weinreich, 1972).
At the same time as features were being introduced into linguistic anal-
yses, the role of a predicate’s valence in relation to syntactic expressibility
began to be studied. Valence, a term introduced by Tesnière (1959), is a char-
acterization of the number and semantic nature of arguments carried by
a verb or other predicative expressions. In the late 1960s and early 1970s,
alternatives to the Katz and Fodor model began to emerge that incorpo-
rated many of the characteristics and principles of valence-based grammars.
These theories attempt to respect the relational structure of sentence mean-
ing while encoding the named “semantic functions” of the arguments in
the lexical entries for predicates (Lakoff, 1965/1970; Fillmore, 1968a, 1969;
Gruber, 1976; Jackendoff, 1972).
Fillmore (1968a, 1977a), for example, uses an enriched notion of valence
to account for how arguments bind to syntactic positions in the sentence.
From these early accounts of case grammar, Fillmore and colleagues devel-
oped a broader notion of frame semantics (cf. Fillmore, 1976, 1982), where
human activities are conceptualized as lexically encoded frames. A seman-
tic frame specifies the conditions under which a predicate combines with its
possible arguments, seen as the participants in the event which that predi-
cate denotes.
Some of these ideas were incorporated into lexically rich, feature-based
semantics, in an attempt to explain how the semantic properties of pred-
icates predict syntactic expressibility and behavior. One version of this
grew into the framework known as generative semantics (Lakoff, 1965/1970;
McCawley, 1968a), where the input to semantic interpretation was the deep
structure of a sentence. While this started as an attempt to explain the
selectional preferences imposed by verbs on their arguments, the scope of
the theory expanded to account for all semantic interpretation from deep
structure.
This view changed with Chomsky’s and Jackendoff’s lexicalist work in the
1970s, where the role of the lexicon became more central to grammatical
processes, and generalizations could be made in terms of what properties
were shared by lexical items (Chomsky, 1970; Jackendoff, 1972). While the
Aspects-model of selectional features restricted the relation of selection to
that between lexical items, work by McCawley (1968a) and Jackendoff (1972)
showed that selectional restrictions must be available to computations at
the level of derived semantic representation rather than at deep structure.
Later work by Bresnan (1982), Gazdar et al. (1985), and Pollard and Sag (1994)
extended the range of phenomena that can be handled by the projection
and exploitation of lexically derived information in the grammar. In these
frameworks, the lexicon plays a central part in the way compositional pro-
cesses are carried out in the grammar.
Before the mid-twentieth century, there was little interest in word mean-
ing within traditional philosophy. While linguistic semantics can trace its
roots back to both Frege (1892b) and Russell (1905), these and other authors
were less interested in word meaning and linguistic behavior than they
were in how words were used as the medium through which judgments can
be formed and inferences made. Frege’s focus lay in formulating the rules
which create meaningful expressions in a compositional manner, while also
introducing an important distinction between an expression’s sense and
its reference. However, because of the descriptive bias in linguistics at the
time, linguists largely failed to appreciate the role that systematic models
of compositionality might play in language generally. Not until mid-century,
with Montague’s synthesis of grammatical description and intensional type
theory, were these issues addressed in a comprehensive (and influential)
manner.
Montague (1970a,b) introduces a bold new program for semantic inter-
pretation in natural language, based on formal logic with a model-theoretic
interpretation. Some of the most influential contemporary work on lexical
semantics is based on this foundation, as we shall see in the sections below.
In Dowty (1979), a model-theoretic interpretation of the decompositional
techniques introduced by Lakoff, McCawley, and Ross was developed.
Dowty’s work, together with Partee (1975)’s seminal article on how
Montague semantics was compatible with syntactic models familiar to

linguists at the time, helped introduce formally grounded approaches to
semantics to the mainstream of the field.
Over the past two decades, there have been serious efforts to create a syn-
thesis of lexical semantics and compositional mechanisms within linguistic
theory. What has emerged is a new view of the role played by the lexicon
regarding how composition is directed. Further, there is a new understand-
ing of how contextualized aspects of interpretation impact the design deci-
sions for what semantics is attributed to lexical items.
Examples of this approach can be seen in varieties of Generative Lexicon
Theory (Pustejovsky, 1995; Bouillon and Busa, 2001), and in work by Jack-
endoff (1997, 2002). These developments have helped to characterize the
approaches to lexical design in terms of a hierarchy of semantic expressive-
ness. There are at least three such classes of lexical description, defined as
follows: Sense Enumerative Lexicons, where lexical items have a single type and
meaning, and ambiguity is treated by multiple listings of words; Polymorphic
Lexicons, where lexical items are active objects, contributing to the deter-
mination of meaning in context, under well-defined constraints; and Unre-
stricted Sense Lexicons, where the meanings of lexical items are determined
mostly by context and conventional use. It seems clear that the most promis-
ing directions seem to be a careful and formal elucidation of polymorphic
lexicons, and this will form the basis of our subsequent discussion.
2.3 Issues in lexical semantics
In this section, we look more closely at a set of phenomena in language that

are of particular relevance to how lexical meaning is encoded and applied
through compositional mechanisms in the grammar. We focus on four spe-
cific problems here: (a) the semantic distinctions between lexical ambiguity
and systematic polysemy; (b) the nature of selectional preferences in verb-
argument composition; (c) the polymorphic behavior of argument selection
by predicates; and (d) the question of lexical representation, more generally
construed.
2.3.1 Ambiguity versus polysemy

Given the compactness of a lexicon relative to the number of objects and
relations referred to in the world, lexical ambiguity would seem inevitable
for any language. Furthermore, the cultural, historical, and linguistic blend-
ing that contributes to the meanings of our words tends to make lexical
ambiguity – when a word carries more than one meaning – appear arbitrary
as well. Hence, homonymy – where one lexical form has many meanings – is
to be expected in a language. Examples of homonyms are illustrated in the
sentences below:
(1) a. Mary strolled along the bank of the river.

b. This company is the largest bank in the city.
(2) a. The doctor treated the woman with antibiotics.

b. He always treats his guests well.
(3) a. First we leave the gate, then we taxi down the runway.
b. John saw the taxi on the street.
(4) a. The judge asked the defendant to approach the bar.

b. The defendant was in the pub at the bar.
The senses of each noun in the examples above are arguably unrelated to
each other; such lexical distinctions have also been called contrastive ambi-
guities (cf. Weinreich, 1972). For this reason, it is assumed that homonyms
are represented as separate lexical entries within the organization of
the lexicon. Words with multiple senses are simply listed separately in
the lexicon, but this does not seem to compromise or complicate the
compositional process of how words combine in the interpretation of a
sentence.
We can compare this to the phenomenon known as polysemy (cf. Apresjan,
1973). Polysemy is the relationship that exists between related senses of a
word, rather than arbitrary ones, as in the above examples. For example,
the noun book is polysemous and can refer to either a physical object or the
information contained in it, as illustrated in (5) below.
(5) a. Mary carried the book home.

b. Mary doesn’t agree with the book.
Unlike the homonyms above, these two senses are logically related to each
other by the very concept of book. Similarly, the noun lunch can refer to the
food intended for consumption at a meal or the actual meal itself, as seen
in (6).
(6) a. Mary has her lunch in her backpack.

b. The phone rang during my lunch yesterday.
In (7), a similar logical relation exists between the two senses of flight; in (7a),
it refers to the event of flying, while in (7b) it refers to the plane engaged in
flying.
(7) a. The flight lasted three hours.

b. The flight landed on time in Los Angeles.
While the two senses of the noun bank in (1), as well as of the other nouns
in examples (2)–(4), are not related (except by phonological form), each
of these examples indicates a formal relation between the lexical senses.
It is the role of the lexicon to distinguish such ambiguities, and to establish
what this logical relation is.
Polysemy is not an isolated phenomenon in language, but rather is asso-

ciated with every major category in the lexicon, within all languages. For
example, adjectives such as good, dangerous, and fast can be viewed as polyse-
mous, where the sense is modulated slightly, depending on the noun being
modified.
(8) a. John is a a good teacher.

b. A good meal is what we need now.
c. Mary took a good umbrella with her into the rain.
In each of these sentences, good is a manner modifier whose interpretation

is dependent on the noun it modifies; in (8a) it means “to teach well”, in
(8b), it means a “tasty meal”, and in (8c), it means “something keeping you
dry”. Similar remarks hold for the adjective dangerous.
(9) a. This is a dangerous road at night.

b. She used a dangerous knife for the turkey.
That is, the road is dangerous in (9a) when “one drives on it”, and the knife
is dangerous in (9b) when “one cuts with it”. Finally, the adjective fast in
the sentences below acts as though it is an adverb, modifying an activity
implicit in the noun; that is, programming in (10a) and driving in (10b).
(10) a. Mary is the fastest programmer we have on staff.

b. The turnpike is a faster road than Main Street.
The exact nature of how adjectives are interpreted relative to the head is
discussed in Pustejovsky (1995), and Busa (1996).
A somewhat related phenomenon involving adjectival scope instead of
true polysemy can be seen in the sentences in (11), where the adjective can
modify the agentive nominal activity or the nominal as an individual.
(11) a. Mary is a beautiful dancer.

b. Jones is a good judge.
This problem has been studied by several researchers (Landman, 1989b;

Pustejovsky, 1995; Busa, 1996; Larson, 1998), and while not involving poly-
semy, it does highlight the need for a potentially richer semantic content of
agentive nominals than is typically assumed.
As mentioned above, polysemy is a phenomenon exhibited by all the
major categories in language, including verbs. Because verbal polysemy is
arguably the most common form of systematic sense relatedness, we cover
it in more detail below.
What is important to realize from the data shown here, for both nouns
and adjectives, is that simply listing the senses of these words will not
always account for their creative use in novel contexts in language. Account-
ing for the behavior of polysemy in language is one of the most difficult
problems facing a theory of the lexicon, and we take up the solutions to
this problem in Section 2.8.
2.3.2 Selectional constraints

To demonstrate the interaction of the lexicon with the other components of
grammar, let us examine how lexical information is accessed and exploited
by syntactic and semantic operations, a process generally referred to as
selection.
By far, the most widely studied type of selection process involves the con-
straints imposed on neighboring phrases by a word, by virtue of its lexical
properties. One of the most important properties of a verb is an encoding
of what phrases it can appear with in the language. In general, this is the
problem of determining how many arguments a verb can appear with in
the syntax and is referred to as argument selection. There is a general rule
of thumb that the number of arguments that the predicate allows in the
syntax corresponds to the number of participants an event has.1 That is,
the argument structure of a word loosely reflects the underlying relational
meaning associated with that word.
For example, consider the behavior of one argument (intransitive), two
argument (transitive), and three argument (ditransitive) verbs in English.
The arity (number of arguments) information is encoded in the verb’s
argument structure. The verbs laugh, see, and give are simple examples in
English.
(12) a. The man laughed.

b. The girl saw a bird.
c. The boy gave a treat to the dog.
The argument structure for the verbs in these sentences can be represented
as follows:
(13) a. laugh(arg1 )
b. see(arg1 ,arg2 )
c. give(arg1 ,arg2 ,arg3 )
The lexicon plays an important role in determining whether a linguistic

expression in a language is well-formed, and selection is the mechanism
through which this is accomplished. For example, it is because of the argu-
ment structures in (13) that these verbs do not appear in the wrong gram-
matical contexts, such as (14) below.
(14) a. *The man laughed the ball.

b. *A bird saw.
c. *The boy gave to the dog.
One of the most important themes when studying the lexicon is to establish
just what role lexical information plays in determining whether an expres-
sion is well-formed or not.
1
We will see below that this is not always the case, however, and there are more complex relations between
semantic representations and the syntactic structures that may appear.
Although the argument structure indicates how many arguments to

expect with a verb in the syntax, it says nothing about what kind of argu-
ments these should be. This can be accomplished by adding selectional con-
straints to the arguments of the verb. A selectional constraint is a require-
ment on an argument that must be satisfied for the verb to be interpreted
properly. The most important constraint imposed on an argument by a pred-
icate is its syntactic category, that is, whether it is a Noun Phrase (NP), Verb
Phrase (VP), Adjective Phrase (AP), or sentence (S). For example, both write
and think are transitive verbs and have the argument structure in (15):
(15) V(arg1 ,arg2 )
However, their arguments have different syntactic requirements, as we can

easily see in (16).
(16) a. The girl bought [ NP some books].

b. The girl thinks [ S her brother is funny].
This can be encoded as part of the argument structure for that verb if we
make a list of features for each argument. In our representation, this list
will be shown directly following the argument it is associated with. So, for
example, let cat be the feature for the category of the argument; then we
have the following distinct argument structures for these two verbs.
(17) a. buy(arg1 [cat=NP], arg2 [cat=NP])

b. think(arg1 [cat=NP], arg2 [cat=S])
Lexical information impacting the grammar Thus far, the lexical informa-
tion needed to interface to the syntax is of two sorts:
r Argument structure of a verb: this determines the number of phrases in the

syntactic construction associated with the verb.
r Syntactic category of each argument: this identifies the actual syntactic
phrase associated with the argument.
There are other selectional phenomena captured by constraints. For exam-

ple, observe the differences in the sentences below.
(18) a. The man / *the rock laughed.

b. The man / the rock fell.
The constraints imposed on the subject by a verb such as laugh would

include the feature of animacy, while a verb like fall does not. Since this
is a lexical property of these verbs, their argument structures should reflect
this distinction. Hence, the verb laugh has a selectional constraint indicating
that the subject must be animate, as shown in (19).
(19) laugh(arg1 [cat=NP,animacy=+])
Similarly, the feature of number is important for selecting the appropriate

subject for verbs such as assemble, gather, and disperse; namely, the subject
must be plural.
(20) a. The students gathered in the courtyard.

b. The children assembled in the gym.
c. The fans dispersed from the stadium.
Notice that it can be either grammatically plural, as in (20) above, or semanti-

cally plural, as with the subjects in (21).
(21) a. The class gathered in the courtyard.

b. The team assembled in the gym.
c. The crowd dispersed from the stadium.
That is, the nouns class, team, and crowd can all be considered as non-atomic,
from the point of view of the verb and the grammar. Hence, the argument
structure for these verbs must reflect this phenomenon by adding a selec-
tional constraint referring to number, ranging over the values “singular”
and “plural”.
(22) assemble(arg1 [cat=NP,num=plural])
It is interesting to ask whether these selectional constraints are indepen-

dent or whether they interact with one another. Let us look a little more
carefully at the verbs in (20) and (21) above. We saw that the subject for each
of these verbs must be plural. How does the feature number interact with
animacy? Notice that both disperse and gather allow non-human subjects,
while assemble does not.
(23) a. The clouds gathered in the sky.

b. *The clouds assembled.
c. The clouds dispersed quickly.
To ensure that a structure such as (23b) is not licensed by the grammar,

there must be an encoding of how the structural conditions are being vio-
lated. Notice that this can already be done using the constraints we’ve dis-
cussed, namely the combination of number and animacy, as illustrated in
(24) below.
(24) assemble(arg1 [cat=NP,num=plural,animacy=+])
It would appear, then, that the verb assemble requires the plural subject to
be animate, while the very similar verbs gather and disperse do not.
Selectional constraints, in fact, can determine the acceptability of argu-
ments in any positions in the grammar; consider the distinction between
the verbs force and convince. Although they are synonyms in many contexts,
force and convince have different selectional properties, the latter requiring
that the “convinced” be a cognitive agent.
(25) a. John forced the guest to leave.

b. John convinced the guest to leave.
(26) a. John forced the door to open.

b. *John convinced the door to open.
Verbs select for different spatial prepositions as well; verbs involving spatial
manner descriptions, for example, are quite specific for the kinds of preposi-
tions they allow. Consider, for example, on and in versus over in the sentences
below, when selected by the verb lie.
(27) The cat is lying on the floor / in the box / *over the floor / . . .
Selectional constraints can also be enlisted to determine the appropriate

selection of manner adverbials for verbs, as well. Notice that in (28), without
some sort of adverbial phrase associated with the verbs behave and perform,
the sentences below are ungrammatical.2
(28) a. Mary behaved *(badly).

b. John performed *(admirably).
We see from this brief discussion that selection is an important part of how
lexical information is conveyed to the syntactic operations in the grammar.
Lexical information impacting the grammar From what we discussed,

there are at least three types of lexical information which interface to the
syntax:
r Argument structure of a verb: this determines the number of phrases in the
syntactic construction associated with the verb.
r Syntactic category of each argument: this identifies the actual syntactic
phrase associated with the argument.
r Selectional constraints of each argument: these identify specific grammat-
ical and semantic features of the argument being selected by the verb.
Selectional constraints have been used to encode all manner of lexical and
syntactic restrictions in grammar. As such, they are part of the meaning
of words and have direct effects on the syntax, morphology, and seman-
tics of the language.
2.3.3 Verb meaning and mapping to syntax

We saw in the previous section the role that the lexicon plays in ensuring
that a linguistic expression is well-formed. Recall that it is the argument
structure that prevents verbs from ending up in the wrong syntactic con-
texts in sentences (Levin, 1993). That is, if the lexicon says it is so, then the
grammar follows.
2
It should be noted that (28a) does have a default reading of “behave well”, when no adverb is present.
(29) a. laugh(arg1 ) ⇒ The man laughed.

b. see(arg1 ,arg2 ) ⇒ The girl saw a bird.
c. give(arg1 ,arg2 ,arg3 ) ⇒ The boy gave a treat to the dog.
Or does it? Notice that there are contexts for each of the verbs in (29), which
exploit properties of the verb, giving rise to constructions not allowed by
the original argument structure.
(30) a. The man laughed himself sick.

b. The girl saw a bird fly into the room.
c. The man gave at the office.
In each of these sentences, the argument structure for the verb has been
violated in some way. In (30a), there is an additional NP object to the verb
laugh, and a predicate modifying the object NP himself. In (30b), what the
girl saw was not just a bird, but what the bird did; namely, “flying into the
room”, described as an additional VP. Finally, in (30c), both of the expected
arguments are missing, and presumably inferred from the context of the
utterance.
These illustrate one aspect of the phenomenon of verbal polysemy. Poly-
semy, as defined in the previous section, is the term given to an ambiguity
where the different meanings of the word are logically related to each other.
Many verbs can appear in multiple contexts taking a different number of
arguments in each, a phenomenon known as an alternation. For example,
the verbs break, roll, and sink all have intransitive and transitive forms, as
shown in the sentences below.
(31) a. The glass broke.

b. Mary broke the glass.
(32) a. The ball rolled down the slide.

b. The boy rolled the ball down the slide.
(33) a. The ship sank.

b. The torpedo sank the ship.
How does the lexicon represent such ambiguities? The simplest way would
be to list the different argument structures for each verb, as shown in (34).
(34) a. break1 (arg1 ); break2 (arg1 ,arg2 )

b. roll1 (arg1 ); roll2 (arg1 ,arg2 )
c. sink1 (arg1 ); sink2 (arg1 ,arg2 )
Note that the semantic role of the intransitive subject is the same as the
transitive object NP in each one of these verbs. That is, it is the undergoer or
patient.
This kind of alternation does not apply to all intransitive verbs, of course,
and the lexicon must somehow prevent verbs like arrive and die from becom-
ing transitive.
(35) a. Your package arrived.

b. *The mailman arrived your package.
(36) a. My computer died yesterday.

b. *The storm died my computer yesterday.
The question arising from such cases is this: what allows for an alternation
to occur for some verbs while not for others? Is this part of the lexicon or
some other rule or strategy in the grammar?
Verb alternations are quite prevalent in language and pose a difficult prob-
lem to lexicon designers. For example, how many different argument struc-
tures do we have for a verb like sweep?
(37) a. John swept.

b. John swept the floor.
c. John swept the dirt into the corner.
d. John swept the dirt off the sidewalk.
e. John swept the floor clean.
f. John swept the dirt into a pile.
Such alternating patterns typically apply to more than just one verb; hence,
it would be more efficient to have a general strategy for how these word
senses are related rather than simply listing the different senses for each
verb.
Lexical information is sensitive to both the syntax and the discourse con-
text. Although a verb may be lexically specified to have a certain number
of arguments, there are many situations in which this can be violated. We
saw this above in (30c) with the verb give. In fact, this is quite common but
appears to be governed by systematic rules, and not just pragmatic informa-
tion. For example, while (30c) is grammatical, for most speakers (38b) below
is not.
(38) a. The woman donated her car to the foundation.

b. *The woman donated.
Similarly, eat and drink can appear without their direct objects, but devour
and gulp cannot (at least for most speakers).
(39) a. The girl ate her lunch quickly.

b. The girl ate quickly.
(40) a. The dog devoured the bone.

b. *The dog devoured.
(41) a. The boy drank his milk quietly.

b. The boy drank quietly.
(42) a. The boy gulped his milk quietly.

b. *The boy gulped quietly.
This type of alternation is typically called indefinite NP deletion and is related

to other “pragmatic deletions”, such as those shown below.
(43) a. John tried to call his mother yesterday.

b. John tried yesterday.
(44) a. John attempted to call his mother yesterday.

b. *John attempted yesterday.
The ability to ignore part of the argument structure in the syntax seems to
be a lexical property, one which is idiosyncratic to each verb.
Finally, we consider a kind of verbal polysemy not involving argument
alternation, but a syntactic category alternation. Recall from our discussion
above (see (17)) that we motivated a way to distinguish between different
syntactic types for an argument; that is, buy and think have different cate-
gory values associated with their second arguments:
(45) a. buy(arg1 [cat=NP],arg2 [cat=NP])

b. think(arg1 [cat=NP],arg2 [cat=S])
What happens, however, when the syntactic distinction involves the same
verb? Consider the sentences below, where the verb begin appears in three
distinct syntactic contexts.
(46) a. Mary began to read the novel.

b. Mary began reading the novel.
c. Mary began the novel.
Verbs like hate and love in English can appear in even more contexts, as seen
in (47) below.
(47) a. John would hate Bill to leave.

b. John hates (it) that Bill left.
c. John hated to lose the game.
d. John hated losing the game.
e. John hated that he lost the game.
The examples above in (46) and (47) bring up the issue of how to make gen-
eralizations in the lexicon; that is, is there a “compact” manner in which to
express that the verb begin means the same thing in each sentence in (46),
and likewise for the verb hate in (47).
2.4 Event semantics
In this section, we look at the notion of event in lexical semantics. There are
two traditions to examine when studying the role of events in the semantics
of language:
Event typology: Predicates in language have an event variable that can be

treated as a first-order individual in the semantics, to enable logical
inference (Davidson, 1967a).
Aktionsarten: Predicates in language can be classified according to their
event type or aspectual class, in order to specifically capture grammati-
cal and semantic behaviors (Vendler, 1957).
The move by Davidson to introduce a first-order event variable in the rep-

resentation was mainly motivated by the need to provide a coherent analysis
of adverbial modification in the interpretation of sentences. Under this pro-
posal, two-place predicates such as eat and three-place predicates such as give
contain an additional argument, the event variable, e, as depicted below.
(48) a. λyλxλe[eat(e, x, y)]

b. λzλyλxλe[give(e, x, y, z)]
In this manner, Davidson is able to capture the appropriate entailments

between propositions involving action and event expressions through the
conventional mechanisms of logical entailment. For example, to capture the
entailments between (49b–d) and (49a) below,
(49) a. Mary ate the soup.

b. Mary ate the soup with a spoon.
c. Mary ate the soup with a spoon in the kitchen.
d. Mary ate the soup with a spoon in the kitchen at 3:00 pm.
In this example, each more specifically described event entails the one
above it by virtue of and-elimination (conjunctive generalization) on the
expression.
(50) a. ∃e[eat(e, m, the_soup)]

b. ∃e[eat(e, m, the_soup) ∧ with(e, a_spoon)]
c. ∃e[eat(e, m, the_soup) ∧ with(e, a_spoon) ∧ in(e, the_kitchen)]
d. ∃e[eat(e, m, the_soup) ∧ with(e, a_spoon) ∧
in(e, the_kitchen) ∧ at(e, 3:00 pm)]
There are of course many variants of the introduction of events into predica-
tive forms, including the identification of arguments with specific named
roles (or partial functions, see Chierchia, 1989; Dowty, 1989).
In lexical semantic analysis, it is standard practice to create component-
based classifications using linguistic data that demonstrate pairwise distinc-
tions for grammatical or semantic well-formedness judgments. One such
approach is the determination of aspectual class or Aktionsart (see also Roth-
stein, Chapter 12). This is essentially a characterization of the different kinds
of eventualities that predicative expressions denote. There have been sev-
eral influential distinctions proposed in the literature, but the best known
are those introduced by Kenny (1963) and Vendler (1967). Kenny assumed
that there are three basic aspectual types: states, activities, and performances.
Vendler proposes a similar distinction for states and processes, but splits
the last class (his events) into two categories, accomplishments and achievements.
His classification as well as his terminology have been the starting point for
much of the work in aspect and event semantics in the field. These event
classes are summarized briefly below.
State: An unchanging situation that holds at a time or over an interval,

with no endpoints.
Activity: A dynamic event with no endpoints (an atelic event).
Accomplishment: An incrementally changing event with a culminating end-
point (a telic event).
Achievement: An instantaneous change of state.
Examples of states are seen in simple attributive predications, such as

(51a–b), as well as with non-dynamic relations, such as (51c–d).
(51) a. Mary is happy.

b. Marc is Dutch.
c. Jan loves Bill.
d. Mary believes Jan is happy.
States can be distinguished from activities by virtue of certain grammatical

diagnostics, many of which Vendler introduced into the literature. Consider
a verb such as swim in sentence (52a). This denotes an activity of unspeci-
fied duration, and the sentence does not convey information regarding a
culmination of this activity. With the addition of a goal to-PP, however, as in
sentence (52b), the swimming activity is bounded, and the resulting event
denotes an accomplishment.
(52) a. Jan swam yesterday.

b. Jan swam to the dock yesterday.
c. Jan is swimming.
Hence, we can analyze the verb swim as lexically denoting an activity,

which can be interpreted contextually as an accomplishment, through the
appropriate compositional modification. Finally, notice that sentence (52c)
denotes a “snapshot” of the swimming process. Both of these aspectual coer-
cions are known to distinguish processes from states.
In a similar fashion, there are verbs which lexically denote accomplish-
ments, such as build and destroy. These verbs encode the logical culmination
to the activity performed during the event, as illustrated in (53) below.
(53) a. The children built a fort.

b. John destroyed the anthill.
Finally, an achievement is an event that results in a change of state,

just as an accomplishment does, but where the change is thought of as
occurring instantaneously. For example, in sentences (54a–c) the change is

not a gradual one, but something that has a point-like quality to it. Hence,
modification by point adverbials such as at 3 pm is suggestive that a sentence
denotes an achievement (see Dowty, 1979).
(54) a. The plane crashed at noon.

b. John found his wallet at 3 pm.
c. The train arrived at midnight.
Pustejovsky (1988, 1991b) extends the decompositional approach presented

in Dowty (1979) by explicitly reifying the events and subevents in the pred-
icative expressions. Unlike Dowty’s treatment of lexical semantics, where
the decompositional calculus builds on propositional or predicative units
(as discussed above), a “syntax of event structure” makes explicit reference
to quantified events as part of the word meaning. Pustejovsky further intro-
duces a tree structure to represent the temporal ordering and dominance
constraints on an event and its subevents (see also Moens and Steedman,
1988).
(55) a. event → state | process | transition

b. state: → e
c. process: → e1 . . . en
d. transitionach : → state state
e. transitionacc : → process state
For example, the accomplishment denoted by “building a house” con-

sists of the building process, followed by the state representing the result
of the object being built. Grimshaw (1990) adopts this theory in her work on
argument structure, where complex events such as break are given a similar
representation. In such structures, the process consists of what an agent,
x, does to cause the breaking of y, and the following state represents the
resulting broken item. The process corresponds to the outer causing event
as discussed above, and the state corresponds in part to the inner change-of-
state event. Both Pustejovsky and Grimshaw differ from the authors above
in assuming a specific level of representation for event structure, distinct
from the representation of other lexical properties. Furthermore, they fol-
low Davidson (1967a), Higginbotham (1985), and particularly Parsons (1990),
in adopting an explicit reference to (at least) one event variable in the param-
eter structure of the verbal semantics.
Recently, Rappaport-Hovav and Levin (2001) and Levin and Rappaport-
Hovav (2005) have adopted a large component of the event structure model
for their analysis of the resultative construction in English; event decom-
position has also been employed for properties of adjectival selection, the
interpretation of compounds, and stage and individual-level predication
(Carlson, 1977a; Diesing, 1992a; Jäger, 2001).
Research done by Tenny (1989), Dowty (1991), Krifka (1992, 1998), and oth-
ers enriches this typology by developing a theory of how an event is shaped
by the incremental participation of the theme in that event. The central

aspect of this view is that an accomplishment is defined as involving a homo-
morphism from parts of the event to parts of the incremental theme. The
problem of incrementality can be illustrated with the following examples.
(56) a. John ate a hamburger.

b. Mary wrote a novel.
The process of eating something is an incremental activity and results in

an accomplishment, described by reference to the quantized unit appearing
in the direct object (theme) position, only when the entire hamburger has
been eaten or the entire novel has been written.
Recent work on scalar change (see Hay et al., 1999; Beavers, 2008; Kennedy
and Levin, 2008) and dynamic event semantics (Naumann, 2001) suggests a
new understanding of the interplay between verb meaning, event seman-
tics, and argument structure with these predicates, by focusing on the mea-
surement of the change in value over the properties of the participants in
each intermediate state during the event.
Starting from a quite different perspective, Krifka (1989a, 1992) presents
an interpretation of Aktionsart using a lattice-theoretic interpretation of
event structure. Using the sum operation from lattice theory, where a com-
plex event e1 e2 can be formed from any two events, e1 and e2 , Krifka intro-
duces a part/whole relation, e1 e2 , iff e2 = e1 e2 . This allows for a distinc-
tion to be made between quantized and cumulative predicates. Processes of
undifferentiated activity, such as walking or singing, are cumulative and are
closed under the sum operator, while an accomplishment, such as build a
house, is not.
(57) a. ∀P[CUM(P) ↔ ∀x∀y[[P(x) ∧ P(y)] → P(x y)]]

b. ∀P[QUA(P) ↔ ∀x∀y[[P(x) ∧ P(y)] → x y]]
These two classes map fairly closely to the well-known categories of atelic
and telic predicates, mentioned above. For example, no part of building a
house is a complete house-building; the event is quantized by the relation-
ship to the result, i.e., the house (see Krifka, 1992). The activity of walking,
on the other hand, is cumulative, and any event of walking is composed of
subevents of walking. See Champollion and Krifka, Chapter 13 for further
discussion.
2.5 Lexical decomposition
Lexical decomposition is concerned with the internal semantic structure of

lexical items within a lexicon. The focus of lexical decomposition is on how
the lexical items are semantically similar and distinct by virtue of shared
knowledge structures or semantic primitives. While numerous structures
and primitives (e.g., semantic features and semantic markers) have been
proposed, we limit our discussion here to the most currently relevant

theories, all of them structural decomposition models, as introduced above.
The goal of lexical decomposition has traditionally been to provide the
necessary and sufficient conditions for the meaning of every lexical item
in a subject domain or language. In many ways, this goal is similar to that
of the syntactic analysis of sentences in a language. If primitives and struc-
tures are taken as an exhaustive set on top of which all expressions in the
language are expressed, then the meaning of any lexical item in the lan-
guage will have to be derived from these terms.
Perhaps the first significant framework of structural decomposition was
generative semantics (see Lakoff, 1965/1970). It emerged just as Katz and
Fodor’s feature-based decomposition model was shown to be both incom-
plete and inadequate (see Weinreich, 1972) as a model for language seman-
tics. Generative semanticists argued that deep structure was the appropriate
level for semantic structure and that transformational rules operate only on
that structure, without subsequent semantic consequences. Lexical items
are decomposed into a set of abstract components and transformations are
done on those components (see Lakoff, 1965/1970; Ross, 1970). For example,
the lexical item kill is decomposed into the predicate DEAD and two higher-
level predicates CAUSE and BECOME. The deep semantic structure associ-
ated with this verb consists of these semantic components organized into a
labeled bracketing.
While the framework of generative semantics is no longer generally
adopted, many aspects of the theory can be seen in contemporary lexical
theories, owing to some significant generalizations the theory made about
semantic structure in language. For example, the sentences in (58) form a
sort of paradigm for the concept dead, related by the application of abstract
predicates, which systematically change the meaning of the words associ-
ated with it (see Lakoff, 1965/1970; McCawley, 1968b; Carter et al., 1988).
(58) a. John killed Bill.

b. Bill died.
c. Bill is dead.
Assuming that the underlying form for a verb like kill encodes the stative
predicate in (58c) and the relation of causation, generative semanticists
posited representations such as (59) below.
(59) (CAUSE(x, (BECOME(NOT(ALIVE y)))))
Here the predicate CAUSE is represented as a relation between an individual

causer x and an expression involving a change of state in the argument y.
Although there is an intuition that the cause relation involves a causer and
an event, neither Lakoff nor Carter makes this commitment explicitly (a
point we return to below).
In a very influential work, Dowty (1979) presents a model-theoretic inter-
pretation of the ideas from generative semantics, but with one important
difference: Dowty associates complex decompositional expressions such as

the one in (59) with the lexical items directly, without the syntactic machin-
ery of predicate reduction that was required in a generative semantics
derivation from deep to surface structure. Hence, the lexicon in Dowty’s
theory is far richer than normally envisioned.
Another significant contribution made by Dowty’s study of word mean-
ing is his reinterpretation of Vendler’s classification of the Aktionsarten,
mentioned in Section 2.4, in model-theoretic terms. Thus, states, activities,
accomplishments, and achievements are given a formal treatment within
an intensional logic.
Pursuing a similar line of reasoning to the generative semanticists con-
cerning lexical decomposition, Jackendoff (1972) built on the Gruber (1976)
thesis to argue that predicates such as [+cause] and [+inchoative] are encoded
in the meaning of the word itself.3 In Jackendoff’s later work (1983, 1990,
2002), the approach to lexical decomposition makes claims for cognitive rel-
evance that are not important motivations for many researchers. However,
Jackendoff believes in the cognitive primacy of the primitives used within
his system, and the role of these primitives in performing inferences. We
examine Jackendoff’s semantic representation briefly here, as it pertains to
lexical decomposition.
At the core of Jackendoff’s conceptual semantics is an inventory of primi-
tive categories from which lexical meanings are constructed. These include
concepts such as event, state, action, place, path, property, and amount.
Along with these are the formation rules used to combine these categories
into concepts associated with lexemes. To illustrate, consider sentence (60)
below, along with its lexical representation in Jackendoff’s system.
(60) a. John flew the plane from New York to Boston.

b. CAUSE(John, GO(plane, [[path] FROM (IN [Boston]),
TO(IN [New York])]))
Interestingly, many of these predicates are spatial in origin. Jackendoff, like

many others interested in the nature of semantic extension, explores the
idea that the predicates within the spatial domain may be used to analyze
concepts within other semantic domains. The basic idea here is that for dif-
ferent semantic fields such as possession, identification, circumstantial, and
temporal, a verb from the spatial field can acquire a new meaning using the
same primitives because it is being evaluated relative to a new field. Exam-
ples of this include:
(61) a. Possessional: John will take the house and Mary the kids.
b. Identificational: John turned into a great father.
c. Circumstantial: Mary led me to believe she was younger than I.
d. Temporal: The seminar has been moved to Wednesdays.
3
Jackendoff (1972), following Chomsky (1970), was philosophically at odds with the claims made by
generative semantics. The interpretive semantics they defended took semantics to be a process of
interpretation over the derived surface form of the sentence, rather than the deep structure.
Regardless of the decompositional approach adopted, there are many ques-

tions that remain: do its primitives provide a complete and exhaustive
description of the concepts of natural language? How open-ended is the sys-
tem? What is the additional information which distinguishes the meaning
of one lexical item from another? That is, the primitives do allow for gener-
alizations, but what allows for the discrimination of concepts (the function
filled by distinguishers in Katz and Fodor’s theory), and how many features
are necessary to that end?
In the approaches described here, the general approach is to use decompo-
sition as “annotations” that reflect components of the meanings of lexical
items, rather than exhaustively decompose the meanings of words, as was
the goal in the earlier days of generative grammar.
More recent versions of lexical representations inspired by generative
semantics and Jackendoff’s decomposition model can be seen in the Lexical
Relational Structures of Hale and Keyser (1993), where syntactic tree struc-
tures are employed to capture the same elements of causation and change
of state as in the representations of Carter, Jackendoff, and Dowty. The work
of Levin and Rappaport, building on Jackendoff’s Lexical Conceptual Struc-
tures, has likewise been influential in further articulating the internal struc-
ture of verb meanings (see Levin and Rappaport-Hovav, 1995).
Generative Lexicon can be seen as a hybrid theory, incorporating aspects
of Jackendoff’s and Dowty’s work with the event semantics of Davidson and
Parsons. It has taken several decades for Davidson’s observations regard-
ing the role of events in the determination of verb meaning to find their
way convincingly into the major linguistic frameworks. Over the past two
decades, a synthesis has emerged which attempts to model verb mean-
ings as complex predicative structures with richer event structures (see Par-
sons, 1990; Pustejovsky, 1991b, 1995; Levin and Rappaport-Hovav, 1995). This
research has developed the idea that the meaning of a verb can be analyzed
into a structured representation of the event that the verb designates, and
has furthermore contributed to the realization that verbs may have com-
plex, internal event structures.
2.6 Semantic roles
While decomposition aims to define lexical meaning in terms of a word’s

internal features, theories of semantic roles can be seen as partial decompo-
sition models, focusing on articulating the function or role that the argu-
ments of a predicate play in the determination of sentence meaning. Hence,
as mentioned above in Section 2.2, we can view semantic roles as enhanced
argument structure specifications.
Argument structure is the syntactic encoding of the functional behavior
of an expression in the object language. In linguistically motivated models
of lexical semantics, one of the basic semantic distinctions between nouns
and verbs is stated in terms of their selectional behavior in syntax (the
valence properties discussed in Section 2.2). The argument structure for a

word can be seen as the simplest specification of its semantics, indicating
the number and type of parameters associated with the lexical item as a
predicate. For example, the verb sleep can be represented as a predicate tak-
ing one argument, love as taking two arguments, while the verb buy takes
two, three, or four arguments, depending on the context.
(62) a. sleep(x)
b. love(x,y)
c. build(x,y,z)
The first theory of semantic roles within a generative model was Case
Grammar, a semantically oriented grammar developed by Fillmore (1968b)
and others, e.g., Anderson (1977) and Starosta (1988). Case was first used
for the morphological analysis of noun endings in, e.g., German and Rus-
sian. Fillmore showed that these noun endings serve the same purpose as
the positioning of nouns and prepositions in lexical surface structures. Fill-
more (1968b) introduced the notion of case grammar and a case frame, a
predicate containing arguments that are a set of obligatory and optional
cases. Implicit in the theory is that each NP in a sentence can be assigned
only one case and that the cases assigned by a verb can be realized only once
in a sentence. He defined a number of cases including:
1. Agentive (A): an animate perceived instigator of the action identified by

the verb.
2. Instrumental (I): an inanimate force or object causally involved in the
action or state identified by the verb.
3. Dative (D): an animate being affected by the action identified by the verb.
4. Objective (O): the semantically most neutral case.
For example, the verb open requires that its objective role be filled: some-
thing must open or be opened. In (63a), only the objective role is filled.
In (63b) and (63c), both agentive and objective roles are filled. In (63d), the
instrumental and objective roles are filled. In (63e), the agentive, objective,
and instrumental roles are filled.
(63) a. The door opened.

b. Mary opened the door.
c. The door was opened by Mary.
d. The key opened the door.
e. Mary opened the door with the key.
Fillmore noted that different case roles can occupy the same grammatical
function, e.g., the grammatical subject is door in (63a) and (63c) which occu-
pies the objective role, Mary in (63b) and (63c) which has the agentive role,
and key in (63d) and (63e) which has an instrumental role.
Fillmore’s theory attempts to explain how the arguments of a predicate
are assigned to particular syntactic structures and is not concerned with
establishing an independent level of semantic representation. To handle

assignment, Fillmore assumed a principle of a case hierarchy, which allowed
the selection of grammatical subject and object in default configurations.
As an example of case assignment, consider the verb break which assigns
both obligatory and optional cases. Obligatory cases include the objective,
while optional cases include both the agentive and the instrumental
cases.
(64) John broke the window with a rock.
In sentence (64), by default assignment, agentive is assigned to the subject,

objective to the object, and instrumental to the phrase within the preposi-
tional phrase. This is a type of default selection which can be violated if the
verbal morphology indicates a structure such as a passive as in (63c).
Fillmore (1977a) notes some deficiencies and limitations of case grammar.
Although case grammar is useful as a guide to decomposition of verb mean-
ing, the theory does nothing to clarify the nature of sense relations like
antonymy and hyponymy. Case grammar allows the recognition of sentences
as partially synonymous by matching case structures. Because it does not
claim to be a complete model for meaning, case grammar avoids the pitfalls
of the decomposition model of Katz–Fodor; on the other hand, because of
its incompleteness, it is unclear what role such a system of cases might play
in a comprehensive semantic theory.
Part of Jackendoff’s early work on the theory of lexical decomposition
(as discussed above) included a formalization of Gruber (1976)’s more lim-
ited notion of thematic role. In this view, which is quite similar to Fill-
more’s conception of cases, Jackendoff (1972) classifies verbs according to
which thematic (theta) roles they assign to their arguments. Thus, theta the-
ory, as this view has come to be known, is a minimal decomposition of
verb meaning, where features and abstract predicates have been replaced
by named functional roles, as in case grammar. The success of theta theory
has come not from how well it characterizes verb denotations, but in how
“functional” roles interact with other principles of syntax to determine the
well-formedness of a sentence.
Thematic relations are now generally defined as partial semantic func-
tions of the event being denoted by the verb or noun, and behave according
to a predefined calculus of roles relations (e.g., Carlson, 1984; Chierchia,
1989; Dowty, 1989). For example, semantic roles such as agent, theme, and
goal can be used to partially determine the meaning of a predicate when
they are associated with the grammatical arguments to a verb.
(65) a. putagent,theme,location
b. borrowrecipient,theme,source
Thematic roles can be ordered relative to each other in terms of an implica-

tional hierarchy. For example, there is considerable use of a universal sub-
ject hierarchy such as shown below (see Fillmore, 1968b; Comrie, 1981).
(66) agent > recipient/benefactive> theme/patient > instrument >

location
Many linguists have questioned the general explanatory coverage of the-

matic roles, however, and have chosen alternative methods for capturing the
generalizations they promised. Dowty (1991) suggests that theta-role gener-
alizations are best captured by entailments associated with the predicate
itself. A theta-role can then be seen as the set of predicate entailments that
are properties of a particular argument to the verb. Characteristic entail-
ments might be thought of as prototype roles, or proto-roles; this allows for
degrees or shades of meaning associated with the arguments to a predicate.
Others have opted for a more semantically neutral set of labels to assign
to the parameters of a relation, whether it is realized as a verb, noun, or
adjective (see Hobbs et al., 1988).
In the late 1970s, Fillmore began to rethink his views of semantics and
developed a “scenes-and-frames” semantics to overcome the deficiencies and
limitations he had noted with case grammar. The basic ideas of the new
semantics, influenced in part by Minsky (1975), were that “people associate
certain scenes with certain linguistic frames” and that “meanings are rela-
tivized to scenes,” i.e., lexical items or expressions are understood by placing
them within scenes or images in which they have some linguistic functions,
e.g., naming.
In the theory, frames are linguistic entities that represent the meanings
of lexical items. Scenes can not only be visual but also refer to “interper-
sonal transactions, standard scenarios, familiar layouts” (Fillmore, 1977b)
and “body image” (the way we orient and classify linguistic frames such as
up-down and left-right).
Fillmore’s account views scenes-and-frames as offering an alternative to
traditional accounts of lexical decomposition issues like lexical ambigu-
ity, synonymy, semantic fields and selection restrictions (Fillmore, 1975).
Synonymy can be understood as “indistinguishable scenes for which the
frame offers lexical options”. A selection restriction is viewed as the rela-
tion between a frame and a scene. “The selection restriction information
about the use of a word can be stated as a specification of the nature of the
appropriate scene” (Fillmore, 1975, p. 71).
Fillmore furthermore argues that the denotations of lexical items can bet-
ter be characterized by a sort of prototype theory; that is, “instead of the
meaning of a linguistic form being represented in terms of a checklist of
conditions that have to be satisfied in order for the form to be appropriately
or truthfully used, it is held that the understanding of meaning requires,
at least for a great many cases, an appeal to an exemplar or prototype” (Fill-
more, 1975). A prototype is an ideal example of some concept to which other
instances of the same concept bear a strong (but not necessarily total) resem-
blance. For example, a prototypical bird might have a beak, wings, feathers,
and be able to fly, but an ostrich would still be recognized as a bird. The
prototype idea can be seen in Berlin and Kay (1969)’s study of color terms,
in Labov’s work on the boundary criteria between cups and bowls (Labov,
1972b), and in the work of the psychologist Rosch and Mervis (1975).
2.7 Qualia structure
Thus far, we have focused on the lexical information associated with verb
entries. All of the major categories, however, are encoded with syntactic
and semantic feature structures that determine their constructional behav-
ior and subsequent meaning at logical form. In Generative Lexicon Theory
(Pustejovsky, 1991a, 1995), it is assumed that word meaning is structured on
the basis of four generative factors, or qualia roles, that capture how humans
understand objects and relations in the world and provide the minimal
explanation for the linguistic behavior of lexical items (these are inspired
in large part by the Moravcsik (1975, 1981) interpretation of Aristotelian
aitia). These are: the formal role: the basic category that distinguishes the
object within a larger domain; the constitutive role: the relation between
an object and its constituent parts; the telic role: its purpose and function;
and the agentive role: factors involved in the object’s origin or “coming into
being”. Qualia structure is at the core of the generative properties of the lexi-
con, since it provides a general strategy for creating new types. For example,
consider the properties of nouns such as rock and chair. These nouns can
be distinguished on the basis of semantic criteria which classify them in
terms of general categories such as natural_kind, artifact_object.
Although very useful, this is not sufficient to discriminate semantic types
in a way that also accounts for their grammatical behavior. A crucial dis-
tinction between rock and chair concerns the properties which differenti-
ate natural kinds from artifacts: functionality plays a crucial role in the pro-
cess of individuation of artifacts, but not of natural kinds. This is reflected
in grammatical behavior, whereby a good chair, or enjoy the chair are well-
formed expressions reflecting the specific purpose for which an artifact is
designed, but good rock or enjoy a rock are odd out of context, since for rock
the functionality (i.e., telic) is undefined. Exceptions exist when new con-
cepts are referred to, such as when the object is construed relative to a spe-
cific activity, such as in The climber enjoyed that rock; rock itself takes on a new
meaning, by virtue of having telicity associated with it, and this is accom-
plished by integration with the semantics of the subject NP. Although chair
and rock are both physical_object, they differ in their mode of com-
ing into being (i.e., agentive): artifacts are manufactured, rocks develop in
nature. Similarly, a concept such as food or cookie has a physical manifesta-
tion or denotation, but also a functional grounding, pertaining to the activ-
ity of “eating”. These distinct aspects of a category are represented by the
qualia structure for that concept, which provides a coherent structuring for
different dimensions of meaning.
By analyzing the semantics of objects in terms of qualia, the classic
domain of entities from Montague Grammar can be organized as a hierar-
chical system of subtypes (sorts), structured into three broadly defined types:
(67) Natural types: Natural kind concepts grounded in the Formal and
Constitutive qualia roles.
Artifactual types: Concepts grounded in the Telic (purpose or function),
or Agentive (origin) qualia roles.
Complex types: Concepts integrating reference to the relation between
at least two types from the other levels.
Qualia structure has proved to be an expressive representational device

and has been adopted by adherents of many other grammatical frameworks.
For example, Jensen and Vikner (1994) and Partee and Borschev (1998) as
well as Borschev and Partee (2001) appeal to qualia structure in the inter-
pretation of the genitive relation in NPs, while many working on the inter-
pretation of noun compounds have developed qualia-based strategies for
interpretation of noun–noun relations (Busa and Johnston, 1996; Johnston
and Busa, 1999; Jackendoff, 2010). Van Valin Jr. and La Polla (1997) and Van
Valin Jr. (2005) have also adopted qualia roles within several aspects of anal-
yses in Role and Reference Grammar (RRG), where nominal semantics have
required finer-grained representations.
2.8 Type theory and the lexicon
How can lexical semantic information be exploited in the compositional

operations responsible for building larger semantic expressions in lan-
guage? In this section we discuss how the representational strategies and
approaches outlined above can be integrated into a uniform framework, one
that can interface with the mechanisms we are familiar with within com-
positional semantics. Specifically, we present a type-theoretic treatment of
lexical semantic information, as developed in Pustejovsky (2006), Asher and
Pustejovsky (2006), and subsequently in Asher (2011).
The basic semantic knowledge associated with a lexical item can be split
into two categories: its semantic type and its selectional type. While the former
identifies the semantic class that a lexical item belongs to (such as entities,
events, and attributes), the latter class specifies the semantic characteristics
associated with the arguments to a lexical item. From a type-theoretic posi-
tion, such as that adopted within formal semantics since Montague, these
are in fact identical, as the type of a functional expression is determined
by the nature of the typing of its arguments (see Partee, 1975, 1978; Bach,
1984).
Figure 2.1 Fragment of a type hierarchy
Following standard assumptions within type-theoretic semantics (see

Chierchia and McConnell-Ginet, 2000), we assume a vocabulary of atomic
and derived types. For the present discussion, we let the atomic types be:
e the general type of entities; and t the type of truth values. Derived types
consist of all functional types, formed with the functional infix, →; hence, for
any types σ and τ , then σ → τ is a type.
As in Montague Grammar and other standard frameworks, we will take
a lexical entry to consist of a lambda term and a type assignment to the
variables in the term. This will then determine via the standard interpreta-
tion for the lambda term a functional type for the whole expression. Unlike
Montague Grammar, however, Type Composition Logic (TCL) (see Asher and
Pustejovsky, 2006; Asher, 2011) has a much richer system of types reflecting
the information conventionally associated with a word in the Generative
Lexicon approach (its qualia structure), and correspondingly more complex
rules for manipulating these types. Such typing knowledge is typically struc-
tured by means of some sort of inheritance mechanism (Evans and Gazdar,
1989, 1996; Copestake, 1993; Pollard and Sag, 1994; Copestake and Briscoe,
1997), and modeled as a lattice or semi-lattice (Carpenter, 1992). Briscoe et al.
(1993) describe a rich system of types for allowing default mechanisms into
lexical type descriptions. An example of such an inheritance structure is
shown in Figure 2.1.
As presented in the previous section, the qualia structure is a system of
relations that characterizes the semantics of a lexical item, and TCL provides
a way to encode this directly in the type associated with a lexeme. Following
Asher and Pustejovsky (2006), we adopt two additional type constructors, ⊗
and •. The constructor ⊗ introduces a qualia relation, Q, such that if σ and
τ are types, then so is σ ⊗Q τ , for Q. The constructor • introduces a complex
type, such that if σ and τ are types, then so is σ • τ .
To illustrate how a richer lexical semantics is reflected directly in the type
structure, notice that a feature structure denoting the qualia structure for
an expression, α, reduces to the type α : (β ⊗A σ ) ⊗T τ in TCL (where A and T

stand for Agentive and Telic qualia, respectively).
⎡ ⎤
α
⎢ ⎡ ⎤⎥
⎢ ⎥
⎢ FORMAL = β ⎥
(68) ⎢ ⎢ ⎥⎥
⎢ ⎢ ⎥⎥
⎢QS = ⎢ TELIC = τ ⎥⎥
⎣ ⎣ ⎦⎦
AGENTIVE = σ
We can define all three basic types from (67) using the qualia and these
constructors. The first two classes in (67) are defined in terms of qualia; a nat-
ural physical object is a subtype formed from predication with the atomic
type, e. The natural types, N , are those subtypes that are formally structured
as a join semi-lattice, N , , in this domain (Pustejovsky, 2001). We say
that this type occupies the value of the formal quale in the type structure.
The creation of a predicate that, in turn, selects for a natural type follows
conventional functional typing assumptions: for any type τ in the subdo-
main of natural types, τ ∈ N , τ → t will be considered a natural functional
type.
Once we have defined natural type entities, their corresponding func-
tional types are defined. The creation of predicates over the subdomain of
natural types follows conventional functional typing assumptions: for any
type τ in the subdomain of natural types, τ ∈ N , τ → t is a natural func-
tional type. This allows us to define natural predicates such as die or touch,
as: λx:eN [die(x)], λy:eN λx : eN [touch(x, y)].
The second class of types, artifactual types, are defined as any type with an
associated telic type. For example, the concept of a potable liquid would
have the typing, liquid ⊗T drink. Similarly, an artifactual entity such as bread
would be typed as, (phys ⊗A bake) ⊗T eat. The creation of functional types
over the domain of artifactual types is defined as follows: for any type τ
in the domain of artifactual entity types, τ ∈ A, τ → t is an artifactual func-
tional type. For example, λx:eA [break(x)]; λy:eA λx:eN [fix(x, y)].
Finally, complex types are constructed through a type-construction oper-
ation (the dot, •) over the domain of Naturals, Artifactuals, and Complex
Types. Consider the noun book, a complex type denoting both the informa-
tional content and the physical manifestation of that content: phys • info.
Other examples include the nouns lunch, school, and promise. Constructing
functional types over the subdomain of complex types is straightforward:
for any type τ in the domain of complex entity types, τ ∈ C, τ → t is a
complex functional type. Examples include verbs such as read: λy:phys • info
λx:eN [read(x, y)].
One of the advantages of introducing a finer-grained system of types is the
ability to explain the selectional constraints associated with the arguments
to a predicate; that is, die selects for an animate entity; assemble selects for a
semantically plural individual, and so on.
Another advantage of a richer type system, however, is the way that it

facilitates new solutions to semantic analyses that were not previously avail-
able. For example, consider the problem of verbal polysemy, mentioned in
Section 2.3.1 above, where a verb is able to appear in multiple syntactic con-
texts, with complementary or additional arguments. The examples below
illustrate this phenomenon with the verb enjoy, an aspectual verb begin, and
an experiencer causative, wake up.
(69) a. Mary enjoyed the movie.

b. Mary enjoyed watching the movie.
(70) a. Mary began a book.

b. Mary began to read a book.
(71) a. The coffee woke John up.

b. John’s drinking the coffee woke him up.
What these sentence pairs illustrate is the process of type coercion, a seman-
tic operation that converts an argument to the type which is expected by a
predicate, where it would otherwise not be accepted (see Pustejovsky, 1995).
This is an operation in the grammar ensuring that the selectional require-
ments on an argument to a predicate are satisfied by the argument in the
compositional process. The rules of coercion presuppose a typed language
such as that outlined above. By allowing lexical items to coerce their argu-
ments, we obviate the enumeration of multiple entries for different senses
of a word.
The notion that a predicate can specify a particular target type for its
argument is a very useful one and intuitively explains the different syntactic
argument forms for the verbs above. In sentences (69) and (70), noun phrases
and verb phrases appear in the same argument position, somehow satisfying
the type required by the verbs enjoy and begin. Similarly, in sentences (71),
noun phrases of very different semantic classes appear as subject of the verb
wake.
If we analyze the different syntactic occurrences of the above verbs as
separate lexical entries, following the sense enumeration theory outlined
in previous sections, we are unable to capture the underlying relatedness
between these entries; namely, that no matter what the syntactic form of
their arguments, the verbs seem to be interpreting all the phrases as events
of some sort. It is exactly this type of complement selection which type coer-
cion allows in the compositional process.
2.9 Open issues in lexical semantics
Contemporary views on what constitutes lexical knowledge still vary signif-

icantly in formal semantics, artificial intelligence, and computational lin-
guistics. To a large extent, the nature of grammar itself is determined by
what information the lexicon supplies to other linguistic components. Even

with this accepted view, the lexicon has been viewed as a mostly passive
module of grammar. What has changed over the past twenty years is the
amount of semantic information that is associated with lexical items, as
well as what is taken to be a lexical item in the first place. What started as a
static repository of pure stems, with active word formation rules generating
everything from inflected forms to derivational compounds, has developed
into a rich library of lexically based syntactic encodings, generative opera-
tions, and type-theoretic processes.
One of the most challenging questions confronting lexical semantics
researchers is perhaps how to integrate to and interface with discourse
semantics and contextual information. If lexical meaning is context-
sensitive at the sentence level, as suggested in previous sections, then
perhaps lexical semantics should directly encode how words are ambiguous
or underspecified at the level of discourse as well. It has been proposed (Las-
carides and Copestake, 1999; Asher and Lascarides, 2003b; Bos, 2004) that
context-sensitive approaches to both sentence composition and discourse
interpretation should have a common view about meaning, some of the
same formal tools, and some of the same problems. There is not only context
sensitivity in both lexical and discourse meaning but also interdependence.
It has alternatively been suggested that this issue reduces to the question
of how “pragmatic enrichment” is formally handled in the interpretation of
a contextually situated utterance. Some researchers argue that pragmatic
or situational frames encode “unarticulated constituents” that fill in the
gaps in the interpretation of an utterance left by a syntactically impover-
ished sentence (Travis, 1981; Bach, 1994; Sperber and Wilson, 1995; Carston,
2002; Recanati, 2004). Countering this view, Stanley (2002a,b), Stanley and
Szabó (2000), and others claim that any truth-conditional effect of extra-
linguistic context to a sentence can be traced to the logical form of the lexi-
cal items involved. Both theories have data to support their arguments and
there are merits to both positions. Some resolution to this question may in
fact come from computational models of discourse and how lexical knowl-
edge is exploited in context.
Acknowledgment
I would like to thank the editors and the anonymous reviewer for valuable
comments and feedback on this chapter.

Lexical Semantics

Uploaded by

Copyright:

Available Formats

Lexical Semantics

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

Lexical Semantics

Uploaded by

Copyright:

Available Formats

2

As the linguistic phenomena associated with lexical semantics become

Atomic predication: No interface language is provided, and there is direct

Atomic predication refers to non-decompositional approaches to lexi-

In this entry, after a brief history of research in lexical semantics, we

2.2 The history of lexical semantics

Interest in the meaning of words goes back as far as philosophical thought

Montague semantics was compatible with syntactic models familiar to

2.3 Issues in lexical semantics

In this section, we look more closely at a set of phenomena in language that

2.3.1 Ambiguity versus polysemy

(1) a. Mary strolled along the bank of the river.

(2) a. The doctor treated the woman with antibiotics.

(4) a. The judge asked the defendant to approach the bar.

(5) a. Mary carried the book home.

(6) a. Mary has her lunch in her backpack.

(7) a. The ﬂight lasted three hours.

Polysemy is not an isolated phenomenon in language, but rather is asso-

(8) a. John is a a good teacher.

In each of these sentences, good is a manner modiﬁer whose interpretation

(9) a. This is a dangerous road at night.

(10) a. Mary is the fastest programmer we have on staff.

(11) a. Mary is a beautiful dancer.

This problem has been studied by several researchers (Landman, 1989b;

2.3.2 Selectional constraints

(12) a. The man laughed.

The lexicon plays an important role in determining whether a linguistic

(14) a. *The man laughed the ball.

Although the argument structure indicates how many arguments to

(15) V(arg1 ,arg2 )

However, their arguments have different syntactic requirements, as we can

(16) a. The girl bought [ NP some books].

(17) a. buy(arg1 [cat=NP], arg2 [cat=NP])

r Argument structure of a verb: this determines the number of phrases in the

There are other selectional phenomena captured by constraints. For exam-

(18) a. The man / *the rock laughed.

The constraints imposed on the subject by a verb such as laugh would

(19) laugh(arg1 [cat=NP,animacy=+])

Similarly, the feature of number is important for selecting the appropriate

(20) a. The students gathered in the courtyard.

Notice that it can be either grammatically plural, as in (20) above, or semanti-

(21) a. The class gathered in the courtyard.

(22) assemble(arg1 [cat=NP,num=plural])

It is interesting to ask whether these selectional constraints are indepen-

(23) a. The clouds gathered in the sky.

To ensure that a structure such as (23b) is not licensed by the grammar,

(24) assemble(arg1 [cat=NP,num=plural,animacy=+])

(25) a. John forced the guest to leave.

(26) a. John forced the door to open.

Selectional constraints can also be enlisted to determine the appropriate

(28) a. Mary behaved *(badly).

Lexical information impacting the grammar From what we discussed,

2.3.3 Verb meaning and mapping to syntax

(29) a. laugh(arg1 ) ⇒ The man laughed.

(30) a. The man laughed himself sick.

(31) a. The glass broke.

(32) a. The ball rolled down the slide.

(33) a. The ship sank.

(34) a. break1 (arg1 ); break2 (arg1 ,arg2 )

(35) a. Your package arrived.

(29) a. laugh(arg1 ) ⇒ The man laughed.

(57) a. ∀P[CUM(P) ↔ ∀x∀y[[P(x) ∧ P(y)] → P(x y)]]