<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1057">
  <Title>Issues in Relating Syntax and Semantics</Title>
  <Section position="1" start_page="0" end_page="282" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> The design of a natural language understanding system is dependent on a rich and structured knowledge representation. Much recent work in computational and cognitive linguistics has focused on representations for syntactic and morphological knowledge. Similarly, the knowledge representation paradigm has proposed many conceptual tools for the representation of semantic knowledge. Combining the two in one system, a system for representing linguistic knowledge, is the essential next task for computational linguistics. This paper will explore some issues which arise in the foundational design of such a representational system. In using the fundamental building block of the grammatical construction (Fillmore 1987), I will reanalyze traditional problems like the passive construction and dative &amp;quot;movement&amp;quot;, showing how a &amp;quot;construction&amp;quot; solution avoids common problems in moi'e traditional grammatical theories. In addition, I will suggest how use of an inheritance hierarchy can replace other linguistic tools for abstraction like lexieal rules, while others can be captured by appealing to shared semantic structures. I will also touch on the issues of lexicalization in theories of grammar - whether information is redundantly specified in the lexicon.</Paragraph>
    <Paragraph position="1"> 2. Representation of Syntactic and Semantic Knowledge A complete and explicit description of a language, a grammar, may be represented in many ways. Rule-based descriptions, for example, describe knowledge of language as knowledge of a set of rules. In contrast, traditional descriptions in the pre-generative framework assume that knowledge of language can be expressed by a knowledge of structures. Fillmore (1987) proposes a return to the traditional notion of the grammatical construction. Construction grammar aims at describing the grammar of a language directly in terms of a collection of grammatical constructions each of which represents a pairing of a syntactic pattern with a meaning structure...(p. 3) Certainly this approach is standard for the representation of lexical knowledge. But extending this approach to larger constructions (the passive construction, the subject-object construction) means bluring the traditional boundaries between the lexicon and syntactic knowledge, a view also called for in Beeker (1976), and Wilensky and Arens (1980). In the construction view, then, idioms, for example, are simply constructions with less variability than other, more productive constructions.</Paragraph>
    <Paragraph position="2"> Given that a grammar is a large collection of these pairs, how are they to be represented7 Here we turn to insights from the field of knowledge /epresentation. The fundamental metaphor chosen as a foundation for a theory of grammar-meaning correspondences is that of the hierarchical stratified network. The use of inheritance hierarchies for semantic knowledge is well-established. Less common, however, is its use for syntactic knowledge (Although see Flickenger, Pollard and Wasow (1985) for inheritence in the lexicon). Fewer still are systems which have used a structured inheritance network as a representational *Thanks to Nigel Ward and Robert Wilea~ky. This research was sponsored in part by the Dofe~e Advance Research Projects Ageacy (DOD), Arpa Ordex No. 4871, monitored by the Space end Naval Warfare Systems Command lmder Contract No. N00039-84.C-0089.</Paragraph>
    <Paragraph position="3"> formalism for both syntactic and semantic knowledge. Notable exceptions are PSI-KLONE (Bobrow and Webber 1980), as well Jacobs's (1985) ACE/KING system.</Paragraph>
    <Paragraph position="4"> The economy and elegance derived from having a grammar defined by a collection of structures defined with a single formalism is a strong argument for this type of unified representation. But this approach, grammatical constructions represented in a unified knowledge-representation formalism, has a number of additional advantages: Use of a single representation scheme allows the correspom dence between syntax and semantics to be represented in its own tight.</Paragraph>
    <Paragraph position="5"> The two types of knowledge can be used together in natural language understanding or generation, thus aiding the development of truly integrated understanding or generation mechanisms.</Paragraph>
    <Paragraph position="6"> Using the same theoretical mechanisms to account for generalizan'ons in beth syntactic and semantic phenomena encourages the development of unified treatments of linguistic abstraction. Thus, for example, structured inherideg tance or other abstraction mechanisms can be used as a replacement for syntactic roles, lexical redundancy roles, and other earlier linguistic tools to capture generalizations.</Paragraph>
    <Paragraph position="7"> And finally, constructs that were originally motivated for the semantic domain, constructs such as prototypes and category theory can be applied to syntactic knowledge, following Lakoff (1987) and his arguments for cognitive grammar. null The explicit pairing of every syntactic form with a semantic one is extremely useful for such understanding techniques as &amp;quot;rule-by-rule&amp;quot; parsing.</Paragraph>
    <Paragraph position="8"> It is interesting to note that the two main conceptual primitives of this approach, the doubly articulated construction and the semantic network, both arise directly from some of the earliest work in modem linguistics, indeed, both are clearly expressed in Saussure's Cours. A construction is precisely a sign in the sense of Saussure. It consists of a syntactic and a semantic component, the correlates of his signifiant and signifi'e. As for the semantic network, the idea that a theory of meaning could best be expressed by describing the relations among concepts is simply Sauasure's (1966/1916) claim that &amp;quot;...each linguistic term derives its value from its opposition to all the other terms.&amp;quot; (p. 88)</Paragraph>
    <Section position="1" start_page="0" end_page="278" type="sub_section">
      <SectionTitle>
2.1. Domains of Knowledge
</SectionTitle>
      <Paragraph position="0"> I will refer to the two domains in the network which describe syntactic and semantic knowledge simply as tl~ Syntactic Domain and the Semantic Domain. The semantic domain includes categorization of the lexical item's semantics with respect to the prototypical ontological primitives (as an Action, Thing, ete). If the construction is headed by a valence-taking type, such as most Actions, the choice of lexical entry will constrain the choice of case roles for each of the constituents. Each valence-bearing type in the lexicon has a case frame - a set of case roles such as agent, patient, or instrument, which it subcategorizes (Fillmore 1968). Also assueiated with each entry is an image-seheme (Lakoff 1987) such as the COrCI'AIN~R Schema, or the PART-WHOLE Schema. In addition, a lexieal entry is linked to a frame  (itt the sense of Fillmore 1982). A frame describes the background constituents and concepts necessary for framing the lexical-semantic definition of a concept. Thus a lexieal entry is linked to a frame which it evokes in the understanding process, and which helps in associating and organizing concept in the generation process.</Paragraph>
      <Paragraph position="1"> The representation of the subcategorized constituent elements of syntactic patterns at the Syntactic domain is done by grammatical relations such as Subject, Object or Complement.</Paragraph>
      <Paragraph position="2"> Grammatical Mapping links relate the Semantic and Syntactic domains, mapping the actual syntactic realization to each of the semantic constituents of the frame - specifying which constituent maps to the syntactic sul~ject, object, and obliques. Following is a sketch of how some of these elements might be related. The grammatical mapph~g between the semantic entry for &amp;quot;rob&amp;quot; and its syntactic realization have been simplilied for expository purposes - a fuller description will be suggested when the passive construction is discussed.</Paragraph>
      <Paragraph position="3"> 3. Accounting for Some Traditional Problems Above: I sketched an outline of a representation for semantic and grammatiea~l knowledge. In this section I give some suggestions toward a more explicit characterization of the representation. I do this in the context of discussing some traditional linguistic problems which provide insight into and constraints on the design of a representational scheme.</Paragraph>
    </Section>
    <Section position="2" start_page="278" end_page="281" type="sub_section">
      <SectionTitle>
3.1. Passlvq~
</SectionTitle>
      <Paragraph position="0"> The passive construction in English has been analyzed as a transformational rule, a lexieal rule, a change in grammatical relations. In each of these views, an active sentence like (1) and a passive like (2)  (1) Marlowe mixed that perfectly loathsome martini.</Paragraph>
      <Paragraph position="1"> (2) That perfectly loathsome martini was mixed by Marlowe.  are related by some sort of rule. Even in the&amp;quot; lexieal theories such as LFG, the passive and active lexical items are.related by lexical rules which makq, ~ ~eference to grammatical relations and other syntactic notions. Thas what these views have in common is an attempt to capture a generalization in language by appealing only to the syntactic component c~f the language.</Paragraph>
      <Paragraph position="2"> In a construction-type theory, one can appeal to the semantic component ~o capture the requisite generalization. Thus the derived predicate adjective (ie &amp;quot;passivized&amp;quot; verb) is related to the verb it morphologicaily resembles, by sharing the same argument structure and ontological category at the Lexleal.Semantie domain.</Paragraph>
      <Paragraph position="3"> But how exactly does the meaning differ? And at what level do we express the productivity of the passive construction? For example, in a lexical theory like LFG, the passive is a redundancy rule that applies to transitive verbs in the lexicon to produce new lexical entries with passive subcategorization frames. But as Rice (1987) points out, the criteria by which the passive can be used depend on much more information than can be stored in the lexicon. For example, the difference between (3) and (4) is not present in the lexical entry for sleep.  (3) This bed has been slept in again by that flea-bitten dog.</Paragraph>
      <Paragraph position="4"> (4) *The living loom was slept in by Mary.</Paragraph>
      <Paragraph position="5">  In other words, the criteria for what makes an acceptable passive include more lban just the subeategorization information of the verb. Rice gives many such examples, and argues that that the passive con~ stmetions makes use of a notion of transitivity which makes reference to the entire conceptualization that an utterance realizes.</Paragraph>
      <Paragraph position="6"> A definition of the passive construction would have to include some characterization of the following constraints:  SEMANTIC FUNCFION: - Focus on the most affected participant in the scene.</Paragraph>
      <Paragraph position="7"> - Construe the verbal process as a property or modifier of this participant.</Paragraph>
      <Paragraph position="8"> SYNTACTIC FORM: - Realize this participant as the subject.</Paragraph>
      <Paragraph position="9"> - Realize the verb (in the past participle form) as a predicate :adjective.</Paragraph>
      <Paragraph position="10">  Note that &amp;quot;most affected participant&amp;quot; must be defined with respect to the prototypical notion of transitivity. Additionally, the realization of the passive participle as a predicate adjective motivates its use with the copula (or with got - note that adjectives can also use this verb with a change-of-state reading).</Paragraph>
      <Paragraph position="11"> This description of the passive construction has been described in such a way that it can be combined with other constructions in a model of language use. In other words, where redundancy rules in a lexical theory may describe an abstraction which has little to do with a pefforo mance model, here a construction is not present unless it is used. Redundancy-type abstraction can be capture by use of the abstraction hierarchy.</Paragraph>
      <Paragraph position="12"> Thus in generation, for example, the speaker might choose a frame of participants, a verbal predicate wlfieh assigns thematic roles, and choose to focus ou a participant which happens to be assigned a non-agentive role, or rather one which would prototypically be realized as a direct object. Then the passive construction might be chosen as one construction to include in the utterance. Such a model is used by McKeown (1985). She note the use of the passive construction to focus on a semantic patient when answering question about missiles, producing &amp;quot;Missiles are carried by water-going vehicles and aircraft...&amp;quot; (p 78). Similarly, in a parsing system, the presence of the passive con- null stmction would indicate the focus on the subject, while the verbal root would enable reconstruction of the thematic roles assigned to the arguments. null 3.2. Redundant Information in the Lexicon - Passive Lexlcal Entries Although this representation of the passive as a construction which can be combined with other constructions in a model of use is general and elegant, should we still redundantly specify the passivized versions of verbs in the lexicon? This relates to the difficult question of when rules are simply redundancy rules, and when they are necessary for a theory. Two arguments for &amp;quot;compiling out&amp;quot; the lexicon, listing every entry redundantly, are given by Bresnan. The first is to avoid intricate feeding relations among rules. As I show later when I discuss the relation of the Passive and Dative constructions, this can be avoided by describing constructions with the understanding that they are to be embedded in a theory of use. The second is the now classic evidence against the derivational theory of complexity (Fodor, Bever, and GaITeR 1974). The evidence that the understanding process is no longer for passive sentences than for active ones was strong evidence against a traditional transformational model Of course, this can also be evidence against a derivational lexical model, if the passive lexical entry is derived from the active one by productive lexical rules. But this argument does not apply to the construction grammar model, for the passive sentence is not derived in any fashion from the active one, but from an underlying semantic conceptualization, iin the same way as the active one.</Paragraph>
      <Paragraph position="13"> However, there are certainly cases where even the most steadfast proponent of syntactic predictability from semantics must agree that the lexicon must include information that might be redundant. Indeed, as Lakoff (1987) points out &amp;quot;Our individual grammars may not be structured optimally... That is, a speaker may learn certain constructions by rote rather than fitting them into his grammar as a whole.&amp;quot; If such lexicalizing of knowledge is necessary in a theory of grammar, we must have a way of representing it when it does occur. If, for example, the English lexicon does represent the passive predicate &amp;quot;mixed&amp;quot; (as opposed to simply including the passive construction and the semar~tic argument structure for Mixing-Action) how would it be represented? Roughly, using the inhetitence hierarchy to instantiate a new node in * the lexical-semantic domain of the network, and then having it multiply inherit from both Mixing-Action and Passive. Of course each of these concepts, Mixing-Action and Passive would have to be further fleshed out - Passive according to the construction specification sketched above, and Mixing-Action with some sort of image-schematic notions (Lakoff 1987 and Langacker 1987). But then the new concept Passive-Mixing-Actinn would inherit the conceptual structure of both of these.</Paragraph>
      <Paragraph position="14"> For example, the ontological Event concerning Marlowe and that martini, and its two syntactic realizations, can be represented as in Figure 2. For clarity's sake I have represented the two different syntactic realizations on either side of the semantics.</Paragraph>
      <Paragraph position="15"> Note that using structured inheritance networks allows the Subject-Agent mapping for Bivalent Verbs to be inherited from the more general Verb-Action mapping. Future directions for the representation system include the ability to assign preferences to different mappings, allowing as representation of the agency hierarchy, and other non-binary phenomena.</Paragraph>
      <Paragraph position="16"> Talmy (1985) notes that just as the passive is a productive way of forming adjectives from semantic Actions, there are also many lexicalized adjectival forms of verbs, forms not derived from such a productive construction. Here there is no choice but to use the lexicalized version of the representation, as a general productive mapping does not exist. Figure 3 shows an example of how such a lexiealized representation might be characterized. In section 3.3 I will discuss the representation of similar such non-productive derived forms, non-productive  derived nominals.</Paragraph>
      <Paragraph position="17"> 3.3. Dative  The essential issue in the design of a grammatical representation is how the representation is to be used. In other words, how construcp tions are to be combined in generation, or how they are to be used in understanding. These issues have not tended to receive a lot of attention, but a few famous construction-combining puzzles exist. One such well-known example is the combination of the Passive with the Dative Shift. (or Dative Movement, or Dative Alternation). The dative shift, first noted in the early 60's, involves verbs such as give which seem to allow two different structuring of its complement arguments:  (5) The bartender gave a martini to Bond.</Paragraph>
      <Paragraph position="18"> (6) The bartender gave Bond a martini.</Paragraph>
      <Paragraph position="19">  A similar or related rule produces such sentences as (7) (7) I baked my mother a cake.</Paragraph>
      <Paragraph position="20"> In a lexical grammar, such as LFG, these rules are lexical redundancy rules which apply to some verbs like &amp;quot;give&amp;quot; and produce new lexical entries. In the new entries, the argument which fills the Goal thematic role is realized with the grammatical relation of direct object, while the Theme role takes the Second Object grammatical reiation. However, many recent semantic studies of these Dative and related examples show that there are strong semantic criteria to the application of the Dative constmction~ (Greene 1974, Oehrle 1975, Baker 1979, Langacker 1986). These studies have shown that the felicitous use of the construction involves the Goal or Benefacted argument ending up with some sort of possession of or control over the Theme. Langacker (1986) notes:  (8) I sent a walrus to Antarctica.</Paragraph>
      <Paragraph position="21"> (9) ?I sent Antarctica a walrus.</Paragraph>
      <Paragraph position="22"> (10) I sent the zoo a walrus.</Paragraph>
      <Paragraph position="23"> Here (9) is deviant because it is difficult to place a continent in the role of a possessor.</Paragraph>
      <Paragraph position="24"> (11) I clear the floor for Bill.</Paragraph>
      <Paragraph position="25"> (12) ?I cleared Bill the floor.</Paragraph>
      <Paragraph position="26"> (13) I cleared Bill a place to sleep on the floor.</Paragraph>
      <Paragraph position="27">  Similarly, (13) makes it much easier to conceive of the area of the floor coming under Bill's control than does (12).</Paragraph>
      <Paragraph position="28"> The point of these examples is similar to the one made by Rice (1987) and noted above. The semantic constraints on when a Dative Direct Object can be used can not be expressed in the lexicon - the difference between (12) and (13) depends on a constmal of the entire utterance conceptualization. Thus the Dative Direct Object construction is stated as a construction in the grammar whose combination with specific lexieal items comes up only in use. A rough characterization of some aspects of this construction might be:  SEMANTIC FUNCTION: - Focus on the affectedness of the Goal (or Beneficiary).</Paragraph>
      <Paragraph position="29"> - Do so by emphasizing the final control or possession the Goal has over the Theme.</Paragraph>
      <Paragraph position="30"> SYNTACTIC FORM: - Realize the Goal as the Direct Object - Realize the Theme as the Second Object.</Paragraph>
      <Paragraph position="31">  Given the idea of this semantic constraint, and our representation of the passive constmetinn, we are ready to explain a more complex problem, that of utterances with both Dative Alternation and Passive constructions, such as (14) or (15). Of course, in the transformation paradigm such examples were handled by rule ordering - if the Dative Movement rule moving &amp;quot;Bond&amp;quot; to object position took place before the Passive role, a structure like (15) resulted, otherwise one like (14).  (15) A martini was given to Bond by file bartender.</Paragraph>
      <Paragraph position="32">  tlowever in a construction-based fmmalism, the interaction between the constructions can be predicted by examining the semantics of the constructions. In tile Dative rule, the Goal participant must be focused on as having some sort of control over the Theme. The point is that what i~ means for a Goal participant to be affected in a Goaltransfer scenalio is for the Goal to be transferred possession or control of tile Theme.</Paragraph>
      <Paragraph position="33"> For different types of scenarios (or Idealized Cognitive Model (Lakoff 1987)), then, transitivity means different things. Rice (1987) takes note of this when she proposes three cognitive domains with separate tra,lsitive event prototypes: the physical, mental, and social domains.</Paragraph>
      <Paragraph position="34"> For example, for Locative-transfer scenarios, what it means to be affected by the action is somewhat different. As many writers have noted (anlong them Anderson 1971 and Foley and Van Valin 1984), in the (b) sentences the locative arguments (the truck and the wall) can be understood as completely affected by the verbal action. In other words, the truck completely filled with hay, the wall completely covered with paint. The (a) sentences do not have this reading.</Paragraph>
      <Paragraph position="35"> For example,  (16a) Bill loaded hay on the truck.</Paragraph>
      <Paragraph position="36"> (16b)Bill loaded file track with hay.</Paragraph>
      <Paragraph position="37"> (17a) Harry sprayed paint on the wall.</Paragraph>
      <Paragraph position="38">  (17b)Harry sprayed the wall with paint.</Paragraph>
      <Paragraph position="39"> Thus &amp;quot;affected&amp;quot; here means something about completion. Of course, this is not surprising. Both Rice (1987) and Hopper and Thompson (1980) do list telic actions as more prototypicaUy transitive than atelic ones. But the important point is that what it means to be affected by an action is different for different action-scenarios. Given some such theory of affectedness in transfer-scenarios, we  can say that the Dative construction is used to mark the affectedness of the Goal in this way. But an affected object is exactly the sot~ of'participant the the Passive is us~ tor focusing on. Thus sentences like (14) and (15) can be characterized simply by whether or not they use certain constructions. Sentence (15) is not lormed by syntactic feeding rules, but by the combination in language use of these two constructions.</Paragraph>
    </Section>
    <Section position="3" start_page="281" end_page="282" type="sub_section">
      <SectionTitle>
3.4. Nominalizations
</SectionTitle>
      <Paragraph position="0"> The correspondence between syntax and semantics is encoded by mapping relations which link groups of syntactic m~d semantic cortstituents in the network. To take a lexical example, the semantic primitive 'idog&amp;quot; might be related to the syntactic entity which instantiates the noun &amp;quot;dog&amp;quot; by one of these relations. Likewise, more complex con-structions are mapped between the syntax and semantics. So, for example, the syntactic construction \[Adj Noun\] might be related to a numher of semantic constructions, specifying the various prototypical ways of combining the semantic inforlnation associated with adjectives and nouns. A complete tbeory of this correspondence would involve showing how every syntactic construction was the realization of some sem,'mtic object. But clearly fire appropriate syntactic construction to begin out investigations is the lexical category. Note that a lexical category is a grammatical construction like any other, differing only in the very frequency of its use.</Paragraph>
      <Paragraph position="1"> It has long been observed that a rough correspondence can be drawn between lcxical categories and an ontological partitioning of the workl. Everyone is familiar with the traditional grammarians' definition of a noun as a word representing a &amp;quot;person, place or tiring&amp;quot;. Naturally, the exceptions to this simplistic analysis are abundant. For example, nominalizations such as destruction secm to be more related to actions than to &amp;quot;things.&amp;quot; But the many exceptions to this analysis caused the real semantic nature of lexical categories to receive less attention than it deserved. As Miller and Johnson-Laird (1976) note, &amp;quot;perhaps \[traditional grammarians\] did not really mean that everything labeled by a noun is a concrete object; perhaps they me~mt that when you use a noun to label something, you tend to conceptualize it as if it were a concrete object (p 527)&amp;quot;. In the spirit of this Whorfian idea that lexical categories reflect a set of ontological categories, the Lexical..</Paragraph>
      <Paragraph position="2"> Semantic Domain includes a set of such ontological primitives as Events, Actions, and Things. Mapping relations in the network relate these primitives to syntactic constructions. In prototypical cases, like the noun &amp;quot;table&amp;quot;, what is semantically a Thing is mapped to a consti~ tuent that is syntactically a noun (or more properly a noun phrase).</Paragraph>
      <Paragraph position="3"> Thus the lexical categories of language are the prototypical realizations of these underlying sernantic regularities. These &amp;quot;prototypical realizations&amp;quot; are simply a version of what Lakoff (1987) called &amp;quot;central principles&amp;quot;. Note that they am also similar to Chomsky's (1986) &amp;quot;Canonical Structural Realizations&amp;quot; of semantic concepts.</Paragraph>
      <Paragraph position="4"> However, as each kind of ontological category can be mapped to different syntactic category, it is also possible to have cases of non-prototypical mapping. Thus what have traditionally been called productive nominalizafions are non-prototypical mappings between Actions and different types of nouns. A nominalization like John's destroying the city is related semantically to an an Action, but is mapped in form to a noun phrase. The Action &amp;quot;destroy&amp;quot; maps into the noun &amp;quot;destroying&amp;quot;. Similarly, remember that the passive construction involved the reanalysis of an Action as a a predicative adjective.</Paragraph>
      <Paragraph position="5"> As for noun-phrase mappings, as Ross (1973) showed, there are a number of noun phrase constructions which are bear different relations to the prototypical noun phrase. He discusses eight different types of noun phrases which he arranges along a Nouniness Squish. These NP's are that-clauses, for NP to V, embedded questions, gerunds, possessive gerunds, action nominals, derived nominals, and Nouns. An cxaminao tion of the syntactic category Noun-Phrase is essential to any discussion of relations between Actions and Nouns (1 use Action here only as an example - of course other ontological categories like States axe  relevant as well). One of the great advantages to a unified representa~ lion for linguistic and conceptual knowledge is that such models as Lakoff's (1987) eatego~3, theories for semantics are automatically applicable to syntactic knowledge.</Paragraph>
      <Paragraph position="6"> But here I will only discuss a few of the lexical issues ilwolved in nominalizations, and in particular the difference between what Ross called &amp;quot;action&amp;quot; and &amp;quot;derived&amp;quot; nominals, or the productive and non-. productive nominalizations.</Paragraph>
      <Paragraph position="7"> For example, the verb &amp;quot;destroy&amp;quot; has two lexical nominalizations the productive or gerundive &amp;quot;destroying&amp;quot;, and the derived or non-productive &amp;quot;destruction&amp;quot;. The nominal use of the gerundive &amp;quot;destroying&amp;quot; is productive in the sense that if a new verb entered the language, say &amp;quot;to xerox&amp;quot;, the native speaker would automatically be able to speak about &amp;quot;xeroxing&amp;quot;, Chomsky (1970) proposed that the relationship of these two types of nominals to their corresponding verb be captmed in two diflerent ways. The relationship between productive noniinalizations and verl)s was captured by means of a transformational rule. Non-productive nominalizations (like &amp;quot;destruction&amp;quot;) are related to the verb by sharing similar phrase struetu~es. To accomplish this, Chomsky pn)posed the X-Bar Convention, the use of generalized cross-categorical phrase structure rules.</Paragraph>
      <Paragraph position="8"> In the network fl'amework, the generalization between verbs and productive nominalizations is handled by structure sharing - they shale descriptions of argument structure in the semantic domain, just as with the passive examples. So a network expressing the relation among &amp;quot;destroy&amp;quot; and &amp;quot;destroying&amp;quot; and their thematic structure would be similar to the passive example in Figure 2.</Paragraph>
      <Paragraph position="9"> Indeed, the abstraction hierarchy allows a way to provide for the creativity of language without explicit rules. Imagine the itwention of a new Action, xerox. Now the concept xerox is inserted under the mo~e general concept Action, and inherits from it the mapping relation to the nominal total. A possible representation is outline in Figure 4.</Paragraph>
      <Paragraph position="10"> Note that this method does not involve a &amp;quot;rule&amp;quot; to capture its gen-eralizations. This may call into question the classic Chomskyan argument from creativity to the need for rule systems.</Paragraph>
      <Paragraph position="11"> In this way we can handle non-productive nominalizations as well as productive ones. A characteristic feature of these non-productive nominalizations is that tbeir semantics is rarely predictable tom the semantics of the underlying verb - they tend to differ in idiosyncratic ways. From a diachronic element, we would say that the non-productive nominalizations have undergone ~mantic drill The net.</Paragraph>
      <Paragraph position="12"> - work representation of the synchronic results of semantic drift are simo ply the creation of a new node which inherits fi'om Thi~Jg at the Semantic level. But note that although this new node may differ in various elements of its semantics, it will rarely drift far enough to change its thematic structure. Thus the similarity in the argument structure between non-productive nominalizations and verbs is not a syntactic fact, as predicted by Chomsky 1970, but a semantic one. This would predict that non-productive nominalizations which have drifted far enough to have changed their argument structure would also differ syntactically from the related verh, This is indeed found to be the case.</Paragraph>
      <Paragraph position="13"> So, for example although we are perfectly comfm~able in saying (18) or  (19), the meaning of &amp;quot;laughter&amp;quot; has drifted c~xough that it is uncon~brt ~ able to use it with an object in (20).</Paragraph>
      <Paragraph position="14"> (18) Who is stupid enough to laugh at Marlowe? (19) Laughing at M'lrlowe is a dangerous proposition.</Paragraph>
      <Paragraph position="15"> (20) *Laughter at Marlowe is rare.</Paragraph>
      <Paragraph position="16">  A redundant lexicalized representation of the nOnoproductiw: nominalizatlons might look much like the one for adjecti;,cs in Figme  Tim deriva~km of verbs from nouns citrates in a markedly diflerent way from dm derivation of nouns from verbs. As Clark and Clark (19&amp;quot;/9) note, a speaker ill using a denominal verb denotes a situation where &amp;quot;the parfait noun denotes one role ill the situation, mid tile remaining surface arguments of file denominal verb denote other roles in the situation.&amp;quot; A sketcll of a solution would involve view-mappings which relate not,ns to their derived verbs in the Semantic domain. In this domain, a heavily context-dependent process creates a new lexical entry through a metapt!orical mapping fl'om the underlying Thing.</Paragraph>
      <Paragraph position="17"> Thus this phenomenon may be much more process-oriented than tile construction model currently allows for.</Paragraph>
      <Paragraph position="18"> Another il;lpoltant issue is tile ability to model prototype effects and other cogniiive issues ill categmy theory. &amp;quot;file ISA links which I have sketched in file lexiealized nlodels are only an impoverished model of categorization. A solution will probably involve some solution such as adding nmnerical weights to nodes and to links to express typicality weights.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>