<?xml version="1.0" standalone="yes"?>
<Paper uid="J91-4003">
  <Title>The Generative Lexicon</Title>
  <Section position="3" start_page="410" end_page="414" type="metho">
    <SectionTitle>
2. Methods in Lexical Semantics
</SectionTitle>
    <Paragraph position="0"> Given what I have said, let us examine the questions presented above in more detail.</Paragraph>
    <Paragraph position="1"> First, let us turn to the issue of methodology. How can we determine the soundness of our method? Are new techniques available now that have not been adequately explored? Very briefly, one can summarize the most essential techniques assumed by the field, in some way, as follows (see, for example Cruse \[1986\]): * On the basis of categorial distinctions, establish the fundamental differences between the grammatical classes; the typical semantic behavior of a word of category X. For example, verbs typically behave as predicators, nouns as arguments.</Paragraph>
    <Paragraph position="2"> * Find distinctions between elements of the same word class on the basis of collocation and cooccurrence tests. For example, the nouns dog and book partition into different selectional classes because of contexts involving animacy, while the nouns book and literature partition into different selectional classes because of a mass/count distinction.</Paragraph>
    <Paragraph position="3">  * Test for distinctions of a grammatical nature on the basis of diathesis; i.e. alternations that are realized in the syntax. For example, break vs. cut in (1) and (2) below (Fillmore 1968; Lakoff 1970; Hale and Keyser 1986): Example 1 a. The glass broke.</Paragraph>
    <Paragraph position="4"> b. John broke the glass.</Paragraph>
    <Paragraph position="5">  Computational Linguistics Volume 17, Number 4 Example 2 a. *The bread cut.</Paragraph>
    <Paragraph position="6"> b. John cut the bread.</Paragraph>
    <Paragraph position="7"> Such alternations reveal subtle distinctions in the semantic and syntactic behavior of such verbs. The lexical semantic representations of these verbs are distinguishable on the basis of such tests.</Paragraph>
    <Paragraph position="8"> Test for entailments in the word senses of a lexical item, in different grammatical contexts. One can distinguish, for example, between context-free and context-sensitive entailments. When the use of a word always entails a certain proposition, we say that the resulting entailment is not dependent on the syntactic context (cf. Katz and Fodor 1963; Karttunen 1971, 1974; Seuren 1985). This is illustrated in Example 3, where a killing always entails a dying.</Paragraph>
    <Paragraph position="9"> Example 3 a. John killed Bill.</Paragraph>
    <Paragraph position="10"> b. Bill died.</Paragraph>
    <Paragraph position="11"> When the same lexical item may carry different entailments in different contexts, we say that the entailments are sensitive to the syntactic contexts; for example, forget in Example 4, Example 4 a. John forgot that he locked the door.</Paragraph>
    <Paragraph position="12"> b. John forgot to lock the door.</Paragraph>
    <Paragraph position="13"> Example 4a has a factive interpretation of forget that 4b does not carry: in fact, 4b is counterfactive. Other cases of contextual specification involve aspectual verbs such as begin and finish as shown in Example 5. Example 5 a. Mary finished the cigarette.</Paragraph>
    <Paragraph position="14"> b. Mary finished her beer.</Paragraph>
    <Paragraph position="15"> The exact meaning of the verb finish varies depending on the object it selects, assuming for these examples the meanings finish smoking or finish drinking.</Paragraph>
    <Paragraph position="16"> Test for the ambiguity of a word. Distinguish between homonymy and polysemy, (cf. Hirst 1987; Wilks 1975b); that is, from the accidental and logical aspects of ambiguity. For example, the homonymy between the two senses of bank in Example 6 is accidental. 2 Example 6 a. the bank of the river b. the richest bank in the city 2 Cf. Weinreich (1972) distinguishes between contrastive and complementary polysemy, essentially covering this same distinction. See Section 4 for discussion.  James Pustejovsky The Generative Lexicon In contrast, the senses in Example 7 exhibit a polysemy (cf. Weinreich  1972; Lakoff 1987).</Paragraph>
    <Paragraph position="17"> Example 7 a. The bank raised its interest rates yesterday (i.e. the institution). b. The store is next to the new bank (i.e. the building). * Establish what the compositional nature of a lexical item is when applied to other words. For example, alleged vs. female in Example 8. Example 8 a. the alleged suspect b. the female suspect  While female behaves as a simple intersective modifier in 8b, certain modifiers such as alleged in 8a cannot be treated as simple attributes; rather, they create an intensional context for the head they modify. An even more difficult problem for compositionality arises from phrases containing frequency adjectives (cf. Stump 1981), as shown in 8c and 8d. Example 8 c. An occasional sailor walks by on the weekend.</Paragraph>
    <Paragraph position="18"> d. Caution: may contain an occasional pit (notice on a box of prunes). The challenge here is that the adjective doesn't modify the nominal head, but the entire proposition containing it (cf. Partee \[1985\] for discussion). A similar difficulty arises with the interpretation of scalar predicates such as fast in Example 9. Both the scale and the relative interpretation being selected for depends on the noun that the predicate is modifying. Example 9 a. a fast typist: one who types quickly b. a fast car: one which can move quickly c. a fast waltz: one with a fast tempo Such data raise serious questions about the principles of compositionality and how ambiguity should be accounted for by a theory of semantics. This just briefly characterizes some of the techniques that have been useful for arriving at pre-theoretic notions of word meaning. What has changed over the years are not so much the methods themselves as the descriptive details provided by each test. One thing that has changed, however -- and this is significant -- is the way computational lexicography has provided stronger techniques and even new tools for lexical semantics research: see Atkins (1987) for sense discrimination tasks; Amsler  (1985), Atkins et al. (forthcoming) for constructing concept taxonomies; Wilks et al. (1988) for establishing semantic relatedness among word senses; and Boguraev and Pustejovsky (forthcoming) for testing new ideas about semantic representations. 3. Descriptive Adequacy of Existing Representations  Turning now to the question of how current theories compare with the coverage of lexical semantic data, there are two generalizations that should be made. First, the  Computational Linguistics Volume 17, Number 4 taxonomic descriptions that have recently been made of verb classes are far superior to the classifications available twenty years ago (see Levin \[1985\] for review). Using mainly the descriptive vocabulary of Talmy (1975, 1985) and Jackendoff (1983), fine and subtle distinctions are drawn that were not captured in the earlier, primitives-based approach of Schank (1972, 1975) or the frame semantics of Fillmore (1968). As an example of the verb classifications developed by various researchers (and compiled by the MIT Lexicon Project; see Levin \[1985, 1989\]), consider the grammatical alternations in the example sentences below (cf. Dow .ty 1991). Example 10  a. John met Mary.</Paragraph>
    <Paragraph position="19"> b. John and Mary met.</Paragraph>
    <Paragraph position="20"> Example 11 a. A car ran into a truck.</Paragraph>
    <Paragraph position="21"> b. A car and a truck ran into each other.</Paragraph>
    <Paragraph position="22"> Example 12 a. A car ran into a tree.</Paragraph>
    <Paragraph position="23"> b. *A car and a tree ran into each other.</Paragraph>
    <Paragraph position="24">  These three pairs show how the semantics of transitive motion verbs (e.g. run into) is similar in some respects to reciprocal verbs such as meet. The important difference, however, is that the reciprocal interpretation requires that both subject and object be animate or moving; hence 12b is ill-formed. (cf. Levin 1989; Dowty 1991). Another example of how diathesis reveals the underlying semantic differences between verbs is illustrated in Examples 13 and 14 below. A construction called the conative (see Hale and Keyser \[1986\] and Levin \[1985\]) involves adding the preposition at to the verb, changing the verb meaning to an action directed toward an object. Example 13 a. Mary cut the bread.</Paragraph>
    <Paragraph position="25"> b. Mary cut at the bread.</Paragraph>
    <Paragraph position="26"> Example 14 a. Mary broke the bread.</Paragraph>
    <Paragraph position="27"> b. *Mary broke at the bread.</Paragraph>
    <Paragraph position="28"> What these data indicate is that the conative is possible only with verbs of a particular semantic class; namely, verbs that specify the manner of an action that results in a change of state of an object.</Paragraph>
    <Paragraph position="29"> As useful and informative as the research on verb classification is, there is a major shortcoming with this approach. Unlike the theories of Katz and Fodor (1963), Wilks (1975a), and Quillian (1968), there is no general coherent view on what the entire lexicon will look like when semantic structures for other major categories are studied. This can be essential for establishing a globally coherent theory of semantic representation. On the other hand, the semantic distinctions captured by these older theories were often too coarse-grained. It is clear, therefore, that the classifications made by Levin and her colleagues are an important starting point for a serious theory of knowledge representation. I claim that lexical semantics must build upon this research toward  James Pustejovsky The Generative Lexicon constructing a theory of word meaning that is integrated into a linguistic theory, as well as interpreted in a real knowledge representation system.</Paragraph>
  </Section>
  <Section position="4" start_page="414" end_page="415" type="metho">
    <SectionTitle>
4. Explanatory Adequacy of Existing Representations
</SectionTitle>
    <Paragraph position="0"> In this section I turn to the question of whether current theories have changed the way we look at representation and lexicon design. The question here is whether the representations assumed by current theories are adequate to account for the richness of natural language semantics. It should be pointed out here that a theory of lexical meaning will affect the general design of our semantic theory in several ways. If we view the goal of a semantic theory as being able to recursively assign meanings to expressions, accounting for phenomena such as synonymy, antonymy, polysemy, metonymy, etc., then our view of compositionality depends ultimately on what the basic lexical categories of the language denote. Conventional wisdom on this point paints a picture of words behaving as either active functors or passive arguments (Montague 1974). But we will see that if we change the way in which categories can denote, then the form of compositionality itself changes. Therefore, if done correctly, lexical semantics can be a means to reevaluate the very nature of semantic composition in language.</Paragraph>
    <Paragraph position="1"> In what ways could lexical semantics affect the larger methods of composition in semantics? I mentioned above that most of the careful representation work has been done on verb classes. In fact, the semantic weight in both lexical and compositional terms usually falls on the verb. This has obvious consequences for how to treat lexical ambiguity. For example, consider the verb bake in the two sentences below.</Paragraph>
    <Paragraph position="2"> Example 15 a. John baked the potato.</Paragraph>
    <Paragraph position="3"> b. John baked the cake.</Paragraph>
    <Paragraph position="4"> Atkins, Kegl, and Levin (1988) demonstrate that verbs such as bake are systematically ambiguous, with both a change-of-state sense (15a) and a create sense (15b).</Paragraph>
    <Paragraph position="5"> A similar ambiguity exists with verbs that allow the resulative construction, shown in Examples 16 and 17, and discussed in Dowty (1979), Jackendoff (1983), and Levin  and Rapoport (1988).</Paragraph>
    <Paragraph position="6"> Example 16 a. Mary hammered the metal.</Paragraph>
    <Paragraph position="7"> b. Mary hammered the metal flat.</Paragraph>
    <Paragraph position="8"> Example 17 a. John wiped the table.</Paragraph>
    <Paragraph position="9"> b. John wiped the table clean.</Paragraph>
    <Paragraph position="10">  On many views, the verbs in Examples 16 and 17 are ambiguous, related by either a lexical transformation (Levin and Rapoport 1988), or a meaning postulate (Dowty 1979). In fact, given strict requirements on the way that a verb can project its lexical information, the verb run in Example 18 will also have two lexical entries, depending on the syntactic environment it selects (Talmy 1985; Levin and Rappaport 1988).  a. Mary ran to the store yesterday.</Paragraph>
    <Paragraph position="11"> b. Mary ran yesterday.</Paragraph>
    <Paragraph position="12"> These two verbs differ in their semantic representations, where run in 18a means goto-by-means-of-running, while in 18b it means simply move-by-running (cf. Jackendoff 1983).</Paragraph>
    <Paragraph position="13"> The methodology described above for distinguishing word senses is also assumed by those working in more formal frameworks. For example, Dowty (1985) proposes multiple entries for control and raising verbs, and establishes their semantic equivalence with the use of meaning postulates. That is, the verbs in Examples 19 and 20 are lexically distinct but semantically related by rules. 3 Example 19 a. It seems that John likes Mary.</Paragraph>
    <Paragraph position="14"> b. John seems to like Mary.</Paragraph>
    <Paragraph position="15"> Example 20 a. Mary prefers that she come.</Paragraph>
    <Paragraph position="16"> b. Mary prefers to come.</Paragraph>
    <Paragraph position="17"> Given the conventional notions of function application and composition, there is little choice but to treat all of the above cases as polysemous verbs. Yet, something about the systematicity of such ambiguity suggests that a more general and simpler explanation should be possible. By relaxing the conditions on how the meaning of a complex expression is derived from its parts, I will, in fact, propose a very straight-forward explanation for these cases of logical polysemy.</Paragraph>
  </Section>
  <Section position="5" start_page="415" end_page="415" type="metho">
    <SectionTitle>
5. A Framework for Computational Semantics
</SectionTitle>
    <Paragraph position="0"> In this section, I will outline what I think are the basic requirements for a theory of computational semantics. I will present a conservative approach to decomposition, where lexical items are minimally decomposed into structured forms (or templates) rather than sets of features. This will provide us with a generative framework for the composition of lexical meanings, thereby defining the well-formedness conditions for semantic expressions in a language.</Paragraph>
    <Paragraph position="1"> We can distinguish between two distinct approaches to the study of word meaning: primitive-based theories and relation-based theories. Those advocating primitives assume that word meaning can be exhaustively defined in terms of a fixed set of primitive elements (e.g. Wilks 1975a; Katz 1972; Lakoff 1971; Schank 1975). Inferences are made through the primitives into which a word is decomposed. In contrast to this view, a relation-based theory of word meaning claims that there is no need for decomposition into primitives if words (and their concepts) are associated through a network of explicitly defined links (e.g. Quillian 1968; Collins and Quillian 1969; Fodor 1975; Carnap 1956; Brachman 1979). Sometimes referred to as meaning postulates, these links establish any inference between words as an explicit part of a network of word</Paragraph>
  </Section>
  <Section position="6" start_page="415" end_page="419" type="metho">
    <SectionTitle>
3 Both Klein and Sag (1985) and Chomsky (1981) assume, however, that there are reasons for relating these two forms structurally. See below and Pustejovsky (1989a) for details.
</SectionTitle>
    <Paragraph position="0"> James Pustejovsky The Generative Lexicon concepts. 4 What I would like to do is to propose a new way of viewing primitives, looking more at the generative or compositional aspects of lexical semantics, rather than the decomposition into a specified number of primitives.</Paragraph>
    <Paragraph position="1"> Most approaches to lexical semantics making use of primitives can be characterized as using some form of feature-based semantics, since the meaning of a word is essentially decomposable into a set of features (e.g. Katz and Fodor 1963; Katz 1972; Wilks 1975; Schank 1975). Even those theories that rely on some internal structure for word meaning (e.g. Dowty 1979; Fillmore 1985) do not provide a complete characterization for all of the well-formed expressions in the language. Jackendoff (1983) comes closest, but falls short of a comprehensive semantics for all categories in language.</Paragraph>
    <Paragraph position="2"> No existing framework, in my view, provides a method for the decomposition of lexical categories.</Paragraph>
    <Paragraph position="3"> What exactly would a method for lexical decomposition give us? Instead of a taxonomy of the concepts in a language, categorized by sets of features, such a method would tell us the minimal semantic configuration of a lexical item. Furthermore, it should tell us the compositional properties of a word, just as a grammar informs us of the specific syntactic behavior of a certain category. What we are led to, therefore, is a generative theory of word meaning, but one very different from the generative semantics of the 1970s.</Paragraph>
    <Paragraph position="4"> To explain why I am suggesting that lexical decomposition proceed in a generative fashion rather than the traditional exhaustive approach, let me take as a classic example, the word closed as used in Example 21 (see Lakoff 1970).</Paragraph>
    <Paragraph position="5"> Example 21 a. The door is closed.</Paragraph>
    <Paragraph position="6"> b. The door closed.</Paragraph>
    <Paragraph position="7"> c. John closed the door.</Paragraph>
    <Paragraph position="8"> Lakoff (1970), Jackendoff (1972), and others have suggested that the sense in 21c must incorporate something like cause-to-become-not-open for its meaning. Similarly, a verb such as give specifies a transfer from one person to another, e.g., cause-to-have. Most decomposition theories assume a set of primitives and then operate within this set to capture the meanings of all the words in the language. These approaches can be called exhaustive since they assume that with a fixed number of primitives, complete definitions of lexical meaning can be given. In the sentences in 21, for example, close is defined in terms of the negation of a primitive, open. Any method assuming a fixed number of primitives, however, runs into some well-known problems with being able to capture the full expressiveness of natural language.</Paragraph>
    <Paragraph position="9"> These problems are not, however, endemic to all decomposition approaches. I would like to suggest that lexical (and conceptual) decomposition is possible if it is performed generatively. Rather than assuming a fixed set of primitives, let us assume a fixed number of generative devices that can be seen as constructing semantic expressions. 5 Just as a formal language is described more in terms of the productions in the grammar than its accompanying vocabulary, a semantic language is definable by the rules generating the structures for expressions rather than the vocabulary of primitives itself. 6  Computational Linguistics Volume 17, Number 4 How might this be done? Consider the sentences in. Example 21 again. A minimal decomposition on the word closed is that it introduces an opposition of terms: closed and not-closed. For the verbal forms in 21b and 21c, both terms in this opposition are predicated of different subevents denoted by the sentences. In 21a, this opposition is left implicit, since the sentence refers to a single state. Any minimal analysis of the semantics of a lexical item can be termed a generative operation, since it operates on the predicate(s) already literally provided by the word. This type of analysis is essentially Aristotle's principle of opposition (cf. Lloyd 1968), and it will form the basis of one level of representation for a lexical item. The essential opposition denoted by a predicate forms part of what I will call the qualia structure of that lexical item. Briefly, the qualia structure of a word specifies four aspects of its meaning: * the relation between it and its constituent parts; * that which distinguishes it within a larger domain (its physical characteristics); * its purpose and function; * whatever brings it about.</Paragraph>
    <Paragraph position="10"> I will call these aspects of a word's meaning its Constitutive Role, Formal Role, Telic Role, and its Agentive Role, respectively. 7 This minimal semantic distinction is given expressive force when combined with a theory of event types. For example, the predicate in 21a denotes the state of the door being closed. No opposition is expressed by this predicate. In 21b and 21c, however, the opposition is explicitly part of the meaning of the predicate. Both these predicates denote what I will call transitions. The intransitive use of close in 21b makes no mention of the causer, yet the transition from not-closed to closed is still entailed. In 21c, the event that brings about the closed state of the door is made more explicit by specifying the actor involved. These differences constitute what I call the event structure of a lexical item. Both the opposition of predicates and the specification of causation are part of a verb's semantics, and are structurally associated with slots in the event template for the word. As we will see in the next section, there are different inferences associated with each event type, as well as different syntactic behaviors (cf. Grimshaw 1990 and Pustejovsky 1991).</Paragraph>
    <Paragraph position="11"> Because the lexical semantic representation of a word is not an isolated expression, but is in fact linked to the rest of the lexicon, in Section 7, I suggest how the global integration of the semantics for a lexical item is achieved by structured inheritance through the different qualia associated with a word. I call this the lexical inheritance structure for the word.</Paragraph>
    <Paragraph position="12"> Finally, we must realize that part of the meaning of a word is how it translates the underlying semantic representations into expressions that are utilized by the syntax. This is what many have called the argument structure for a lexical item. I will build on Grimshaw's recent proposals (Grimshaw 1990) for how to define the mapping from the lexicon to syntax.</Paragraph>
    <Paragraph position="13"> to a particular vocabulary of primitives, a lexical semantics should provide a method for the decomposition and composition of lexical items.</Paragraph>
    <Paragraph position="14"> 7 Some of these roles are reminiscent of descriptors used by various computational researchers, such as Wilks (1975b), Hayes (1979), and Hobbs et al. (1987). Within the theory outlined here, these roles determine a minimal semantic description of a word that has both semantic and grammatical consequences.</Paragraph>
    <Paragraph position="15">  James Pustejovsky The Generative Lexicon This provides us with an answer to the question of what levels of semantic representation are necessary for a computational lexical semantics. In sum, I will argue  that lexical meaning can best be captured by assuming the following levels of representation. null 1. Argument Structure: The behavior of a word as a function, with its arity specified. This is the predicate argument structure for a word, which indicates how it maps to syntactic expressions.</Paragraph>
    <Paragraph position="16"> 2. Event Structure: Identification of the particular event type (in the sense of Vendler \[1967\]) for a word or phrase: e.g. as state, process, or transition.</Paragraph>
    <Paragraph position="17"> 3. Qualia Structure: The essential attributes of an object as defined by the lexical item.</Paragraph>
    <Paragraph position="18"> 4. Inheritance Structure: How the word is globally related to other  concepts in the lexicon.</Paragraph>
    <Paragraph position="19"> These four structures essentially constitute the different levels of semantic expressiveness and representation that are needed for a computational theory of lexical semantics. Each level contributes a different kind of information to the meaning of a word. The important difference between this highly configurational approach to lexical semantics and feature-based approaches is that the recursive calculus defined for word meaning here also provides the foundation for a fully compositional semantics for natural language and its interpretation into a knowledge representation model.</Paragraph>
    <Section position="1" start_page="418" end_page="418" type="sub_section">
      <SectionTitle>
5.1 Argument Structure
</SectionTitle>
      <Paragraph position="0"> A logical starting point for our investigations into the meaning of words is what has been called the functional structure or argument structure associated with verbs. What originally began as the simple listing of the parameters or arguments associated with a predicate has developed into a sophisticated view of the way arguments are mapped onto syntactic expressions (for example, the f-structure in Lexical Functional Grammar \[Bresnan 1982\] and the Projection Principle in Government-Binding Theory \[Chomsky 1981\]).</Paragraph>
      <Paragraph position="1"> One of the most important contributions has been the view that argument structure is highly structured independent of the syntax. Williams's (1981) distinction between external and internal arguments and Grimshaw's proposal for a hierarchically structured representation (Grimshaw 1990) provide us with the basic syntax for one aspect of a word's meaning.</Paragraph>
      <Paragraph position="2"> The argument structure for a word can be seen as a minimal specification of its lexical semantics. By itself, it is certainly inadequate for capturing the semantic characterization of a lexical item, but it is a necessary component.</Paragraph>
    </Section>
    <Section position="2" start_page="418" end_page="419" type="sub_section">
      <SectionTitle>
5.2 Event Structure
</SectionTitle>
      <Paragraph position="0"> As mentioned above, the theory of decomposition being outlined here is based on the central idea that word meaning is highly structured, and not simply a set of semantic features. Let us assume this is the case. Then the lexical items in a language will essentially be generated by the recursive principles of our semantic theory. One level of semantic description involves an event-based interpretation of a word or phrase.</Paragraph>
      <Paragraph position="1"> I will call this level the event structure of a word (cf. Pustejovsky 1991; Moens and Steedman 1988). The event structure of a word is one level of the semantic specification  Computational Linguistics Volume 17, Number 4 for a lexical item, along with its argument structure, qualia structure, and inheritance structure. Because it is recursively defined on the syntax, it is also a property of phrases and sentences. 8 I will assume a sortal distinction between three classes of events: states (eS), processes (8), and transitions (eT). Unlike most previous ,;ortal classifications for events, I will adopt a subeventual analysis or predicates, as argued in Pustejovsky (1991) and independently proposed in Croft (1991). In this view, an event sort such as e T may be decomposed into two sequentially structured subevents, (ee,sS). Aspects of the proposal will be introduced as needed in the following discussion.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="419" end_page="436" type="metho">
    <SectionTitle>
6. A Theory of Qualia
</SectionTitle>
    <Paragraph position="0"> In Section 5, I demonstrated how most of the lexical semantics research has concentrated on verbal semantics. This bias influences our analyses of how to handle ambiguity and certain noncompositional structures. Therefore, the only way to relate the different senses for the verbs in the examples below was to posit separate entries.</Paragraph>
    <Paragraph position="1"> Example 22 a. John baked the potato.</Paragraph>
    <Paragraph position="3"> a. John wiped the table.</Paragraph>
    <Paragraph position="4"> (wipe1 = change (x, State (y) )) b. John wiped the table clean.</Paragraph>
    <Paragraph position="6"> b. Mary ran to the store yesterday.</Paragraph>
    <Paragraph position="8"> * Although the complement types selected by bake in 22, for example, are semantically related, the two word senses are clearly distinct and therefore must be lexically distinguished. According to the sense enumeration view, the same argument holds for the verbs in 23-25 as well.</Paragraph>
    <Paragraph position="9"> 8 This proposal is an extension of ideas explored by Bach (1986), Higginbotham (1985), and Allen (1984). For a full discussion, see Pustejovsky (1988, 1991). See Tenny (1987) for a proposal on how aspectual distinctions are mapped to the syntax.  James Pustejovsky The Generative Lexicon A similar philosophy has lead linguists to multiply word senses in constructions involving Control and Equi-verbs, where different syntactic contexts necessitate different semantic types? Example 26 a. It seems that John likes Mary.</Paragraph>
    <Paragraph position="10"> b. John seems to like Mary.</Paragraph>
    <Paragraph position="11"> Example 27  a. Mary prefers that she come.</Paragraph>
    <Paragraph position="12"> b. Mary prefers to come.</Paragraph>
    <Paragraph position="13"> Normally, compositionality in such structures simply refers to the application of the functional element, the verb, to its arguments. Yet, such examples indicate that in order to capture the systematicity of such ambiguity, something else is at play, where a richer notion of composition is operative. What then accounts for the polysemy of the verbs in the examples above? The basic idea I will pursue is the following. Rather than treating the expressions that behave as arguments to a function as simple, passive objects, imagine that they are as active in the semantics as the verb itself. The product of function application would be sensitive to both the function and its active argument. Something like this is suggested in Keenan and Faltz (1985), as the Meaning-Form Correlation Principle. I will refer to such behavior as cocompositionality (see below). What I have in mind can best be illustrated by returning to the examples in 28.</Paragraph>
    <Paragraph position="14"> Example 28 a. John baked the potato.</Paragraph>
    <Paragraph position="15"> b. John baked the cake.</Paragraph>
    <Paragraph position="16"> Rather than having two separate word senses for a verb such as bake, suppose there is simply one, a change-of-state reading. Without going into the details of the analysis, let us assume that bake can be lexically specified as denoting a process verb, and is minimally represented as Example 29. ldeg Example 29 Lexical Semantics for bake: n )~y)~x)~eP\[bake(e P) A agent(e P, x) A object(e P, y)\] In order to explain the shift in meaning of the verb, we need to specify more clearly what the lexical semantics of a noun is. I have argued above that lexical semantic theory must make a logical distinction between the following qualia roles: the constitutive, formal, telic, and agentive roles. Now let us examine these roles in more detail. One can distinguish between potato and cake in terms of how they come about; the former 9 For example, Dowty (1985) proposes multiple entries for verbs taking different subcategorizations. Gazdar et al. (1985), adopting the analysis in Klein and Sag (1985), propose a set of lexical type-shifting operations to capture sense relatedness. We return to this topic below. 10 I will be assuming a Davidsonian-style representation for the discussion below. Predicates in the language are typed for a particular event-sort, and thematic roles are treated as partial functions over the event (cf. Dowty 1989 and Chierchia 1989).</Paragraph>
    <Paragraph position="17"> 11 More precisely, the process e p should reflect that it is the substance contained in the object x that is affected. See footnote 20 for explanation.</Paragraph>
    <Paragraph position="18">  Computational Linguistics Volume 17, Number 4 is a natural kind, while the latter is an artifact. Knowledge of an object includes not just being able to identify or refer, but more specifically, being able to explain how an artifact comes into being, as well as what it is used for; the denotation of an object must identify these roles. Thus, any artifact can be identified with the state of being that object, relative to certain predicates.</Paragraph>
    <Paragraph position="19"> As is well known from work on event semantics and Aktionsarten, it is a general property of processes that they can shift their event type to become a transition event (cf. Hinrichs 1985; Moens and Steedman 1987; and Krifka 1987). This particular fact about event structures, together with the semantic distinction made above between the two object types, provides us with an explanation for what I will refer to as the logical polysemy of verbs such as bake.</Paragraph>
    <Paragraph position="20"> As illustrated in Example 30a, when the verb takes as its complement a natural kind such as potato, the resulting semantic interpretation is unchanged; i.e., a process reading of a state-change. This is because the noun does not &amp;quot;project&amp;quot; an event structure of its own. That is, relative to the process of baking, potato does not denote an event-type. 12  from a state-change predicate to its creation sense? As just suggested, this additional meaning is contributed by specific lexical knowledge we have about artifacts, and cake in particular; namely, there is an event associated with that object's &amp;quot;coming into being,&amp;quot; in this case the process of baking. Thus, just as a verb can select for an argument-type, we can imagine that an argument is itself able to select the predicates that govern it. I will refer to such constructions as cospecifications. Informally, relative to the process bake, the noun cake carries the selectional information that it is a process  of &amp;quot;baking&amp;quot; that brings it about. 13 12 However, relative to the process of growing, the noun potato does denote an event: 1. Mary grew the potato.</Paragraph>
    <Paragraph position="21"> 13 Other examples of cospecifications are: a. read a book, b. smoke a cigarette, c. mail a letter, d. deliver a lecture, and e. take a bath.  James Pustejovsky The Generative Lexicon We can illustrate this schematically in Example 31, where the complement effectively acts like a &amp;quot;stage-level&amp;quot; event predicate (cf. Carlson 1977) relative to the process event-type of the verb (i.e. a function from processes to transitions, &lt;P,T&gt;). 14 The change in meaning in 31 comes not from the semantics of bake, but rather in composition with the complement of the verb, at the level of the entire verb phrase. The &amp;quot;creation&amp;quot; sense arises from the semantic role of cake that specifies it is an artifact (see below for discussion).</Paragraph>
    <Paragraph position="22"> Example 31 a. bake as a derived Transition: 15  Thus, we can derive both word senses of verbs like bake by putting some of the semantic weight on the NP. This view suggests that, in such cases, the verb itself is not polysemous. Rather, the sense of &amp;quot;create&amp;quot; is part of the meaning of cake by virtue of it being an artifact. The verb appears polysemous because certain complements add to the basic meaning by virtue of what they denote. We return to this topic below, There are several interesting things about such collocations. First, because the complement &amp;quot;selects&amp;quot; the verb that governs it (by virtue of knowledge of what is done to the object), the semantics of the phrase is changed. The semantic &amp;quot;connectedness,&amp;quot; as it were, is tighter when cospecification obtains. In such cases, the verb is able to successfully drop the dative PP argument, as shown below in (1). When the complement does not select the verb governing it, dative-drop is ungrammatical as seen in (2) (although there are predicates selected by these nouns; e.g. keep a secret, read a book, and play a record). la. Romeo gave the lecture.</Paragraph>
    <Paragraph position="23">  b.Hamlet mailed a letter.</Paragraph>
    <Paragraph position="24"> c. Cordelia told a story.</Paragraph>
    <Paragraph position="25"> d.Gertrude showed a movie.</Paragraph>
    <Paragraph position="26"> e.Mary asked a question.</Paragraph>
    <Paragraph position="27"> 2a. *Bill told the secret.</Paragraph>
    <Paragraph position="28"> b.*Mary gave a book.</Paragraph>
    <Paragraph position="29"> c. *Cordelia showed the record.</Paragraph>
    <Paragraph position="30"> For discussion see Pustejovsky (in press).</Paragraph>
    <Paragraph position="31"> 14 cf. Pustejovsky (forthcoming) for details.</Paragraph>
    <Paragraph position="32"> 15 As mentioned in footnote 11, this representation is incomplete. See footnote 20 for semantics of bake.  Computational Linguistics Volume 17, Number 4 and provide a formal treatment for how the nominal semantics is expressed in these examples.</Paragraph>
    <Paragraph position="33"> Similar principles seem to be operating in tile resultative constructions in Examples 23 and 24; namely, a systematic ambiguity is the result of principles of semantic composition rather than lexical ambiguity of the verbs. For example, the resultative interpretations for the verbs hammer in 23(b) and wipe in 24(b) arise from a similar operation, where both verbs are underlyingly specified with an event type of process. The adjectival phrases fiat and clean, although clearly stative in nature, can also be interpreted as stage-level event predicates (cf. Dowty 1979). Notice, then, how the resultative construction requires no additional word sense for the verb, nor any special semantic machinery for the resultative interpretation to be available. Schematically, this is shown in Example 32.</Paragraph>
    <Paragraph position="35"> John hammer the metal flat</Paragraph>
    <Paragraph position="37"> In fact, this analysis explains why it is that only process verbs participate in the resultative construction, and why the resultant phrase (the adjectival phrase) must be a subset of the states, namely, stage-level event predicates. Because the meaning of the sentence in 32 is determined by both function application of hammer to its arguments and function application of fiat to the event-type of the verb, this is an example of cocompositionality (cf. Pustejovsky \[forthcoming\] for discussion).</Paragraph>
    <Paragraph position="38"> Having discussed some of the behavior of logical polysemy in verbs; let us continue our discussion of lexical ambiguity with the issue of metonymy. Metonymy, where a subpart or related part of an object &amp;quot;stands for&amp;quot; the object itself, also poses a problem for standard denotational theories of semantics. To see why, imagine how our semantics could account for the &amp;quot;reference shifts&amp;quot; of the complements shown in Ex- null a. Mary enjoyed the book.</Paragraph>
    <Paragraph position="39"> b. Thatcher vetoed the channel tunnel. (Cf. Hobbs 1987) c. John began a novel.</Paragraph>
    <Paragraph position="40"> 16 See Nunberg (1978) and Fauconnier (1985) for very clear discussions of the semantics of metonymy and the nature of reference shifts. See Wilks (1975) and Fass (1988) for computational models of metonymic resolution.</Paragraph>
    <Paragraph position="41">  James Pustejovsky The Generative Lexicon The complements of enjoy in 33(a) and begin in 33(c) are not what these verbs normally select for semantically, namely a property or action. Similarly, the verb veto normally selects for an object that is a legislative bill or a suggestion. Syntactically, these may simply be additional subcategorizations, but how are these examples related semantically to the normal interpretations? I suggest that these are cases of semantic type coercion (cf. Pustejovsky 1989a), where the verb has coerced the meaning of a term phrase into a different semantic type. Briefly, type coercion can be defined as follows: 17</Paragraph>
    <Section position="1" start_page="424" end_page="434" type="sub_section">
      <SectionTitle>
Definition
</SectionTitle>
      <Paragraph position="0"> Type Coercion: A semantic operation that converts an argument to the type that is expected by a function, where it would otherwise result in a type error.</Paragraph>
      <Paragraph position="1"> In the case of 33(b), it is obvious that what is vetoed is some proposal relating to the object denoted by the tunnel. In 33(a), the book is enjoyed only by virtue of some event or process that involves the book, performed by Mary. It might furthermore be reasonable to assume that the semantic structure of book specifies what the artifact is used for; i.e. reading. Such a coercion results in a word sense for the NP that I will call logical metonymy. Roughly, logical metonymy occurs when a logical argument (i.e.</Paragraph>
      <Paragraph position="2"> subpart) of a semantic type that is selected by some function denotes the semantic type itself.</Paragraph>
      <Paragraph position="3"> Another interesting set of examples involves the possible subjects of causative verbs. TM Consider the sentences in Examples 34 and 35.</Paragraph>
      <Paragraph position="4"> Example 34 a. Driving a car in Boston frightens me.</Paragraph>
      <Paragraph position="5"> b. To drive a car in Boston frightens me.</Paragraph>
      <Paragraph position="6"> c. Driving frightens me.</Paragraph>
      <Paragraph position="7"> d. John's driving frightens me.</Paragraph>
      <Paragraph position="8"> e. Cars frighten me.</Paragraph>
      <Paragraph position="9"> f. Listening to this music upsets me.</Paragraph>
      <Paragraph position="10"> g. This music upsets me.</Paragraph>
      <Paragraph position="11"> h. To listen to this music would upset me.</Paragraph>
      <Paragraph position="12"> Example 35 a. John killed Mary.</Paragraph>
      <Paragraph position="13"> b. The gun killed Mary.</Paragraph>
      <Paragraph position="14"> c. John's stupidity killed Mary.</Paragraph>
      <Paragraph position="15"> d. The war killed Mary.</Paragraph>
      <Paragraph position="16"> e. John's pulling the trigger killed Mary.</Paragraph>
      <Paragraph position="17">  As these examples illustrate, the syntactic argument to a verb is not always the same logical argument in the semantic relation. Although superficially similar to cases of general metonymy (cf. Lakoff and Johnson 1982; Nunberg 1978), there is an interesting systematicity to such shifts in meaning that we will try to characterize below as logical metonymy.</Paragraph>
      <Paragraph position="18"> 171 am following Cardelli and Wegener (1985) and their characterization of polyrnorphismic behavior. 18 See Verma and Mohanan (1991) for an extensive survey of experiencer subject constructions in different languages.</Paragraph>
      <Paragraph position="19">  Computational Linguistics Volume 17, Number 4 The sentences in 34 illustrate the various syntactic consequences of metonymy and coercion involving experiencer verbs, while those in 35 show the different metonymic extensions possible from the causing event in a killing. The generalization here is that  when a verb selects an event as one of its arguments, type coercion to an event will permit a limited range of logical metonymies. For example, in sentences 34(a,b,c,d,f,h), the ' entire event is directly referred to, while in 34(e,g) only a participant from the coerced event reading is directly expressed. Other examples of coercion include &amp;quot;concealed questions&amp;quot; 36 and &amp;quot;concealed exclamations&amp;quot; 37 (cf. Grimshaw 1979; Elliott 1974). Example 36 a. John knows the plane's arrival time.</Paragraph>
      <Paragraph position="20"> (= what time the plane will arrive) b. Bill figured out the answer.</Paragraph>
      <Paragraph position="21"> (= what the answer is) Example 37 a. John shocked me with his bad behavior.</Paragraph>
      <Paragraph position="22"> (= how bad his behavior is) b. You'd be surprised at the big cars he buys.</Paragraph>
      <Paragraph position="23"> (= how big the cars he buys are) That is, although the italicized phrases syntactically appear as NPs, their semantics is  the same as if the verbs had selected an overt question or exclamation. In explaining the behavior of the systematic ambiguity above, I made reference to properties of the noun phrase that are not typical semantic properties for nouns in linguistics; e.g., artifact, natural kind. In Pustejovsky (1989b) and Pustejovsky and Anick (1988), I suggest that there is a system of relations that characterizes the semantics of nominals, very much like the argument structure of a verb. I called this the Qualia Structure, inspired by Aristotle's theory of explanation and ideas from Moravcsik (1975). Essentially, the qualia structure of a noun determines its meaning as much as the list of arguments determines a verb's meaning. The elements that make up a qualia structure include notions such as container, space, surface, figure, artifact, and so on. 19 As stated earlier, there are four basic roles that constitute the qualia structure for a lexical item. Here I will elaborate on what these roles are and why they are useful. They are given in Example 38, where each role is defined, along with the possible values that these roles may assume.</Paragraph>
      <Paragraph position="24">  1. Constitutive Role: the relation between an object and its constituents, or proper parts.</Paragraph>
      <Paragraph position="25"> * Material * Weight * Parts and component elements 19 These components of an object's denotation have long been considered crucial for our commonsense understanding of how things interact in the world. Cf. Hayes (1979), Hobbs et al. (1987), and Croft (1991) for discussion of these qualitative aspects of meaning.  When we combine the qualia structure of a NP with the argument structure of a verb, we begin to see a richer notion of compositionality emerging, one that looks very much like object-oriented approaches to programming (cf. Ingria and Pustejovsky 1990). To illustrate these structures at play, let us consider a few examples. Assume that the decompositional semantics of a nominal includes a specification of its qualia structure:</Paragraph>
      <Paragraph position="27"> For example, a minimal semantic description for the noun novel will include values for each of these roles, as shown in Example 40, where *x* can be seen as a distinguished variable, representing the object itself.</Paragraph>
      <Paragraph position="28">  This structures our basic knowledge about the object: it is a narrative; typically in the form of a book; for the purpose of reading (whose event type is a transition); and is an artifact created by a transition event of writing. Observe how this structure differs minimally, but significantly, from the qualia structure for the noun dictionary in Example 41.</Paragraph>
      <Paragraph position="29">  Notice the differences in the values for the constitutive and telic roles. The purpose of a dictionary is an activity of referencing, which has an event structure of a process. I will now demonstrate that such structured information is not only useful for nouns, but necessary to account for their semantic behavior. I suggested earlier, that for cases such as 33, repeated below, there was no need to posit a separate lexical entry for each verb, where the syntactic and semantic types had to be represented explicitly. Example 42 a. Mary enjoyed the book.</Paragraph>
      <Paragraph position="30"> b. Thatcher vetoed the channel tunnel.</Paragraph>
      <Paragraph position="31"> c. John began a novel.</Paragraph>
      <Paragraph position="32"> Rather, the verb was analyzed as coercing its complement to the semantic type it expected. To illustrate this, consider 42(c). The type for begin within a standard typed intensional logic is &lt;VP, &lt;NP, S&gt;&gt;, and its lexical semantics is similar to that of other subject control verbs (cf. Klein and Sag \[1985\] for discussion).</Paragraph>
      <Paragraph position="33">  Because the verb requires that its first argument be of type transition the complement in 33(c) will not match without some sort of shift. It is just this kind of context where the complement (in this case a novel) is coerced to another type. The coercion dictates to the complement that it must conform to its type specification and the qualia roles may  20 It should be pointed out that the lexical structure for the verb bake given above in 30 and 31 can more properly be characterized as a process acting on various qualia of the arguments. 1. AyAxAee eS\[bake(e e) A agent(d, x) A object(d, Const(y) ) A cake(e s) A object(e s, Formal(x))\]  James Pustejovsky The Generative Lexicon in fact have values matching the correct type. For purposes of illustration, the qualia structure for novel from 41 can be represented as the logical expression in Example 46.</Paragraph>
      <Paragraph position="35"> The coercion operation on the complement in the above examples can be seen as a request to find any transition event associated with the noun. As we saw above, the qualia structure contains just this kind of information.</Paragraph>
      <Paragraph position="36"> We can imagine the qualia roles as partial functions from a noun denotation into its subconstituent denotations. For our present purposes, we abbreviate these functions as Q~, Qo QT, QA. When applied, they return the value of a particular qualia role. For example, the purpose of a novel is for reading it, shown in 47(a), while the mode of creating a novel is by writing it, represented in 47(b).</Paragraph>
      <Paragraph position="38"> As the expressions in 47 suggest, there are, in fact, two obvious interpretations for this sentence in 42(c).</Paragraph>
      <Paragraph position="39"> Example 48 a. John began to read a novel.</Paragraph>
      <Paragraph position="40"> b. John began to write a novel.</Paragraph>
      <Paragraph position="41"> One of these is selected by the coercing verb, resulting in a complement that has a event-predicate interpretation, without any syntactic transformations (cf. Pustejovsky \[1989a\] for details), al The derivation in 49(a) and the structure in 49(b) show the effects of this coercion on the verb's complement, using the telic value of novel. 22 21 There are, of course, an indefinite number of interpretations, depending on pragmatic factors and various contextual influences. But I maintain that there are only a finite number of default interpretations available in such constructions. These form part of the lexical semantics of the noun. Additional evidence for this distinction is given in Pustejovsky and Anick (1988) and Briscoe et al. (1990).</Paragraph>
      <Paragraph position="42"> 22 Partee and Rooth (1983) suggest that all expressions in the language can be assigned a base type, while also being associated with a type ladder. Pustejovsky (1989a) extends this proposal, and argues that each expression c~ may have available to it, a set of shifting operators, which we call Ga, which operate over an expression, changing its type and denotation. By making reference to these operators directly in the rule of function application, we can treat the functor polymorphically, as illustrated below.  1. Function Application with Coercion (FAc): If c~ is of type (b, a}, and fl is of type c, then (a) if type c = b, then c~(fl) is of type a.</Paragraph>
      <Paragraph position="43"> (b) if there is a C/ C ~,~ such that cr(fl) results in an expression of type b, then c~(a(fl)) is of type a.</Paragraph>
      <Paragraph position="44"> (c) otherwise a type error is produced.</Paragraph>
      <Paragraph position="45">  The fact that this is not a unique interpretation of the elliptical event predicate is in some ways irrelevant to the notion of type coercion. That there is some event involving the complement is required by the lexical semantics of the governing verb and the rules of type well-formedness, and although there are many ways to act on a novel, I argue that certain relations are &amp;quot;privileged&amp;quot; in the lexical semantics of the noun. It is not the role of a lexical semantic theory to say what readings are preferred, but rather which are available. 23 Assuming the semantic selection given above for begin is correct, we would predict that, because of the process event-type associated with the telic role for dictionary, there is only one default interpretation for the sentence in 50; namely, the agentive event of &amp;quot;compiling.&amp;quot; 23 There are interesting differences in complement types between finish and complete. The former takes both NP and a gerundive VP, while the latter takes only an NP (cf. for example, Freed \[1979\] for discussion).</Paragraph>
      <Paragraph position="46"> la. John finished the book.</Paragraph>
      <Paragraph position="47"> b. John finished writing the book.</Paragraph>
      <Paragraph position="48"> 2a. John completed the book.</Paragraph>
      <Paragraph position="49"> b. *John completed writing the book.</Paragraph>
      <Paragraph position="50"> The difference would indicate that, contrary to some views (e.g. Wierzbicka \[1988\] and Dixon \[1991\]), lexical items need to carry both syntactic and semantic selectional information to determine the range of complements they may take. Notice here also that complete tends to select the agentive role value for its complement and not the telic role. The scope of semantic selection is explored at length in Pustejovsky (forthcoming).</Paragraph>
      <Paragraph position="51">  a. Mary began a dictionary. (Agentive) b. ?? Mary began a dictionary. (Telic) Not surprisingly, when the noun in complement position has no default interpretation within an event predicate -- as given by its qualia structure -- the resulting sentence is extremely odd.</Paragraph>
      <Paragraph position="52"> Example 51 a. *Mary began a rock.</Paragraph>
      <Paragraph position="53"> b. ??John finished the flower.</Paragraph>
      <Paragraph position="54"> The semantic distinctions that are possible once we give semantic weight to lexical items other than verbs are quite wide-ranging. The next example I will consider concerns scalar modifiers, such as fast, that modify different predicates depending on the head they modify. If we think of certain modifiers as modifying only a subset of the qualia for a noun, then we can view fast as modifying only the telic role of an object. This allows us to go beyond treating adjectives such as fast as intersective modifiers for example, as Ax\[car'(x) Afast'(x)\]. Let us assume that an adjective such as fast is a member of the general type (N, N), but can be subtyped as applying to the Telic role of the noun being modified. That is, it has as its type, (IN Telic\],N}. This gives rise directly to the different interpretations in Example 52.</Paragraph>
      <Paragraph position="55"> Example 52 a. a fast car: driving</Paragraph>
      <Paragraph position="57"> These interpretations are all derived from a single word sense for fast. Because the lexical semantics for this adjective indicates that it modifies the telic role of the noun, it effectively acts as an event predicate rather than an attribute over the entire noun denotation, as illustrated in Example 53 for fast motorway (cf. Pustejovsky and Boguraev \[1991\] for discussion).</Paragraph>
      <Paragraph position="58"> Example 53 Ax \[mot orway ( x ) . . . \[Tel ic ( x ) = AeV \[ travel (cars ) ( e P) A on(x)(cars)(e p) A fast(e P) \]\]\] As our final example of how the qualia structure contributes to the semantic interpretation of a sentence, observe how the nominals window and door in Examples 54 and 55 carry two interpretations (cf. Lakoff \[1987\] and Pustejovsky and Anick \[1988\]): Example 54 a. John crawled through the window.</Paragraph>
      <Paragraph position="59"> b. The window is closed.</Paragraph>
      <Paragraph position="60">  a. Mary painted the door.</Paragraph>
      <Paragraph position="61"> b. Mary walked through the door.</Paragraph>
      <Paragraph position="62"> Each noun appears to have two word senses: a physical object denotation and an aperture denotation. Pustejovsky and Anick (1988) characterize the meaning of such &amp;quot;Double Figure-Ground&amp;quot; nominals as inherently relational, where both parameters are logically part of the meaning of the noun. In terms of the qualia structure for this class of nouns, the formal role takes as its value the Figure of a physical object, while the constitutive role assumes the Invert-Figure value of an aperture. 24  The foregrounding or backgrounding of a nominal's qualia is very similar to argument structure-changing operations for verbs. That is, in 55(a), paint applies to the formal role of the door, while in 55(b), through will apply to the constitutive interpretation of the same NP. The ambiguity with such nouns is a logical one, one that is intimately linked to the semantic representation of the object itself. The qualia structure, then, is a way of capturing this logical polysemy.</Paragraph>
      <Paragraph position="63"> In conclusion, it should be pointed out that the entire lexicon is organized around such logical ambiguities, which Pustejovsky and Anick (1988) call Lexical Conceptual Paradigms. Pustejovsky (forthcoming) distinguishes tlhe following systems and the paradigms that lexical items fall into:  Such paradigms provide a means for accounting for the systematic ambiguity that may exist for a lexical item. For example, a noun behaving according to paradigm 57(a) 24 There are many such classes of nominals, both two-dimensional such as those mentioned in the text, and three-dimensional, such as &amp;quot;room,&amp;quot; &amp;quot;fireplace,&amp;quot; and &amp;quot;pipe.&amp;quot; They are interesting semantically, because they are logically ambiguous, referring to either the object or the aperture, but not both. Boguraev and Pustejovsky (forthcoming) show how these logical polysemies are in fact encoded in dictionary definitions for these words.  James Pustejovsky The Generative Lexicon exhibits a logical polysemy involving packaging or grinding operators; e.g., haddock or lamb (cf. Copestake and Briscoe \[1991\] for details).</Paragraph>
      <Paragraph position="64"> 7. Lexical Inheritance Theory  In previous sections, I discussed lexical ambiguity and showed how a richer view of lexical semantics allows us to view a word's meaning as being flexible, where word senses could arise generatively by composition with other words. The final aspect of this flexibility deals with the logical associations a word has in a given context; that is, how this semantic information is organized as a global knowledge base. This involves capturing both the inheritance relations between concepts and, just as importantly, how the concepts are integrated into a coherent expression in a given sentence. I will assume that there are two inheritance mechanisms at work for representing the conceptual relations in the lexicon: fixed inheritance and projective inheritance. The first includes the methods of inheritance traditionally assumed in AI and lexical research (e.g. Roberts and Goldstein 1977; Brachman and Schmolze 1985; Bobrow and Winograd 1977); that is, a fixed network of relations, which is traversed to discover existing related and associated concepts (e.g. hyponyms and hypernyms). In order to arrive at a comprehensive theory of the lexicon, we need to address the issue of global organization, and this involves looking at the various modes of inheritance that exist in language and conceptualization. Some of the best work addressing the issue of how the lexical semantics of a word ties into its deeper conceptual structure includes that of Hobbs et al. (1987) and Wilks (1975), while interesting work on shared information structures in NLP domains is that of Flickinger et al. (1985) and Evans and Gazdar (1989, 1990).</Paragraph>
      <Paragraph position="65"> In addition to this static representation, I will introduce another mechanism for structuring lexical knowledge, the projective inheritance, which operates generatively from the qualia structure of a lexical item to create a relational structure for ad hoc categories. Both are necessary for projecting the semantic representations of individual lexical items onto a sentence level interpretation. The discussion here, however, will be limited to a description of projective inheritance and the notion of &amp;quot;degrees of prototypicality&amp;quot; of predication. I will argue that such degrees of salience or coherence relations can be explained in structural terms by examining a network of related lexical items. 25 I will illustrate the distinction between these mechanisms by considering the two sentences in Example 58, and their relative prototypicality.</Paragraph>
      <Paragraph position="66"> Example 58 a. The prisoner escaped last night.</Paragraph>
      <Paragraph position="67"> b. The prisoner ate dinner last night.</Paragraph>
      <Paragraph position="68"> Both of these sentences are obviously well-formed syntactically, but there is a definite sense that the predication in 58(a) is &amp;quot;tighter&amp;quot; or more prototypical than that in 58(b). What would account for such a difference? Intuitively, we associate prisoner with an escaping event more strongly than an eating event. Yet this is not information that comes from a fixed inheritance structure, but is rather usually assumed to be commonsense knowledge. In what follows, however, I will show that such distinctions 25 Anick and Pustejovsky (1990) explore how metrics such as association ratios can be used to statistically measure the notions of prototypicality mentioned here.</Paragraph>
      <Paragraph position="69">  Computational Linguistics Volume 17, Number 4 can be captured within a theory of lexical semantics by means of generating ad hoc categories.</Paragraph>
      <Paragraph position="70"> First, we give a definition for the fixed inheritance structure of a lexical item (cf. Touretzky 1986). Let Q and P be concepts in our model of lexical organization. Then: Definition A sequence (Q1, P1,..., Pn) is an inheritance path, which can be read as the conjunction of ordered pairs {(xl,yi) \[ 1 &lt; i &lt; n}.</Paragraph>
      <Paragraph position="71"> Furthermore, following Touretsky, from this we can define the set of concepts that lie on an inheritance path, the conclusion space.</Paragraph>
      <Paragraph position="72"> Definition The conclusion space of a set of sequences * is the set of all pairs (Q, P) such that a sequence (Q,..., P) appears in q~.</Paragraph>
      <Paragraph position="73"> From these two definitions we can define the traditional is-a relation, relating the above pairs by a generalization operator, ~G, 26 as well as other relations that I will not discuss. 27 Let us suppose that, in addition to these fixed relational structures, our semantics allows us to dynamically create arbitrary concepts through the application of certain transformations to lexical meanings. For example, for any predicate, Q -- e.g. the value of a qualia role -- we can generate its opposition, ~Q (cf. Pustejovsky 1991). By relating these two predicates temporally we can generate the arbitrary transition events for this opposition (cf. Wright 1963):</Paragraph>
      <Paragraph position="75"> Similarly, by operating over other qualia role values we can generate semantically related concepts. I will call any operator that performs such an operation a projective transformation, and define them below: Definition A projective transformation, Tr, on a predicate Q1 generates a predicate, Q2, such that 7r(Q1) = Q2, where Q2 ~ ~. The set of transformations includes: 7, negation, _&lt;, temporal precedence, &gt;_, temporal succession, =, temporal equivalence, and act, an operator adding agency to an argument.</Paragraph>
      <Paragraph position="76"> Intuitively, the space of concepts traversed by the application of such operators will be related expressions in the neighborhood of the original lexical item. This space can be characterized by the following two definitions: 26 See, for example, Michalski (1983) and Smolka (1988) for a treatment making use of subsorts. 27 Such relations include not only hypernymy and hopyonymy, but also troponymy, which relates verbs by manner relations (cf. Miller 1985; Beckwith et al. 1989; Miller and Fellbaum 1991.</Paragraph>
    </Section>
    <Section position="2" start_page="434" end_page="436" type="sub_section">
      <SectionTitle>
James Pustejovsky The Generative Lexicon
Definition
</SectionTitle>
      <Paragraph position="0"> A series of applications of transformations, ~rl,..., 7rn, generates a sequence of predicates, (Q1,..., Qnl, called the projective expansion of Q1, P(Q1).</Paragraph>
      <Paragraph position="1"> Definition The projective conclusion space, P(@R), is the set of projective expansions generated from all elements of the conclusion space, ~, on role R of predicate Q: as: P(~R) = {(P(Q1),P(Qn)&gt; \[ (QI,...,Qn) E ~R}.</Paragraph>
      <Paragraph position="2"> From this resulting representation, we can generate a relational structure that can be considered the set of ad hoc categories and relations associated with a lexical item (cf. Barselou 1983).</Paragraph>
      <Paragraph position="3"> Using these definitions, let us return to the sentences in Example 58. I will assume that the noun prisoner has a qualia structure such as that shown in 60.</Paragraph>
      <Paragraph position="4">  Using the representation in 60 above, I now trace part of the derivation of the projective conclusion space for prisoner. Inheritance structures are defined for each qualia role of an element. In the case above, values are specified for only two roles. For each role, R, we apply a projective transformation 7r onto the predicate Q that is the value of that role. For example, from the telic role of prisoner we can generalize (e.g. drop the conjunct) to the concept of being confined. From this concept, we can apply the negation operator, generating the predicate opposition of not-confined and confined. To this, we apply the two temporal operators, &lt; and &gt;, generating two states: free before capture and free after capture. Finally, to these concepts, if we apply the operator act, varying who is responsible for the resulting transition event, we generate the concepts: turn in, capture, escape, and release.</Paragraph>
      <Paragraph position="5">  a. ~c: \[confine(y, x) A loc(x, prison)\] ~ confine(y, x) b. 7:BE1 \[-~confine(E1, y, x)\] c. 3E2 \[confine(E2, y, x)\] d. ~: E1 _&lt; E2 = T1 e. &lt;_: E2 &lt;_ E1 = T2 f. act: act(x, T1) = &amp;quot;turn in&amp;quot;  Computational Linguistics Volume 17, Number 4 g. act: act(y, T1) = &amp;quot;capture&amp;quot; h. act: act(x, T2) = &amp;quot;escape&amp;quot; i. act: act(y, T2) = &amp;quot;release&amp;quot; These relations constitute the projective conclusion space for the telic role of prisoner relative to the application of the transformations mentioned above. Similar operations on the formal role will generate concepts such as die and kill. Generating such structures for all items in a sentence during analysis, we can take those graphs that result in no contradictions to be the legitimate semantic interpretations of the entire sentence. Let us now return to the sentences in Example 58. It is now clear why these two sentences differ in their prototypicality (or the relevance conditions on their predication). The predicate eat is not within the space of related concepts generated from the semantics of the NP the prisoner; escape, however, did fall within the projective conclusion space for the Telic role of prisoner, as shown in Example 63.  We can therefore use such a procedure as one metric for evaluating the &amp;quot;proximity&amp;quot; of a predication (Quillian 1968; Hobbs 1982). In the examples above, the difference  James Pustejovsky The Generative Lexicon in semanticality can now be seen as a structural distinction between the semantic representations for the elements in the sentence.</Paragraph>
      <Paragraph position="6"> In this section, I have shown how the lexical inheritance structure of an item relates, in a generative fashion, the decompositional structure of a word to a much larger set of concepts that are related in obvious ways. What we have not addressed, however, is how the fixed inheritance information of a lexical item is formally derivable during composition. This issue is explicitly addressed in Briscoe et al. (1990) as well as Pustejovsky and Briscoe (1991).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>