<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2008">
  <Title>Enjoy the Paper: Lexical Semantics via Lexicology</Title>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1. Lexlcal Semantics
</SectionTitle>
    <Paragraph position="0"> A theory of lexical semantics should provide an efficient representation of lexical semantic information in the paradigmatic plane which is capable of integrating with a genuinely compositional semantic account in the syntagmatic plane. Our starting point for this research is the work of Levin (e.g. 1985) and others on verbal alternations (diathesis), Pustejovsky (e.g. 1989) on lexical coercion and qualia theory, and Evans &amp; Gazdar (e.g.</Paragraph>
    <Paragraph position="1"> 1989) on default inheritance within unification-based formalisms. It can be seen as a further contribution to the use of unification-based formahsms in linguistic description and specifically as an enriching of the minimal sort-based lexical semantic taxonomy incorporated into the Esprit ACORD system (Moens et al., 1989) and the SRI (Cambridge) CLE system (Alshawi et al., 1989). We propose a system in which a standard graph-based unification formalism, such as PATR-II, is augmented with minimal disjunction (of atomic terms) and minimal default inheritance (allowing only 'orthogonal' multiple inheritance in a manner similar to Evans &amp; Gazdar's DATR). Using such a system we are able to see the beginnings of solutions to three problems concerning the integration of lexical semantics with a general theory of linguistic description and processing - alternations, coercion, and decomposition / representation.</Paragraph>
    <Paragraph position="2"> The first problem emerges with systems, such as the Alvey Tools grammar (Carroll &amp; Grover, 1989), which attempt to characterise the grammatical behaviour of lexical items in terms of sets of subcategorisation frames. Intuitively, this often seems arbitrary and inelegant because the occurrence of alternation seems to be semantically motivated. This problem has been discussed in connection with w~rbs mostly, but also arises with nouns and adjectiw'~a. For instance, in the Tools lexicon the verb believe has eight entries. Six of these separate entries relate to the same or a very similar sense of believe; namely, believe3 (Longman Dictionary of Contemporary English, LDOCE) 'to hold as an opinion; suppose' which is a relatlon between an individual (the believer) and a proposition (what is believed). Treating the various grammatical realisations of this sense of believe separately predicts that it is pure accident that they share the same sense. It also suggests that the range of possible alternations is unpredictable and must simply be listed from verb to verb. Most of the work on alternations has concentrated on attempts to characterise semantic classes of verbs which undergo similar alternations (e.g. Levin, 1985). This enterprise has not been particularly successful (Boguraev &amp; Briscoe, 1989b), but in any case ignores or simply assumes the prior point that it is possible to construct a system in which there is just one entry for believe3.</Paragraph>
    <Paragraph position="3"> Nevertheless, it seems correct that examples like John believed that Mary was clever / Mary (to be) clever / Mary / the rumour should be related to one entry for believe because this would allow us to account for the interpretation of John believed Mary as something like 'John believed something(s) that Mary asserted'; thai: is, as standing for some 'understood' proposition involving Mary. Pustejovsky (1989b) refers to this process as coercion and compares it to examples such as John considers Mary a genius where it is usual (e.g. in GPSG, Gazdar et al., 1985) to claim that a genius functions predicatively because the subcategorisation frame for consider forces this interpretation. In general, coercion is a problem in theories which take the syntactic aspect of grammatical realisation as primary, but would be a natural consequence of a theory which took the sense and rite fact that believe3 is a relation between an individual and a proposition as basic. In such an account an NP complement of a verb denoting a relation between an individual and a proposition would either denote a proposition 'directly' (the rumour) or be coerced to the appropriate semantic type (Mary).</Paragraph>
    <Paragraph position="4"> When coercion occurs some additional information is required to 'flesh out' the elevated semantic type of the complement. Pustejovsky (1989) dubs this logical metonymy. In the case of believed Mary this is that it is some assertion of Mary's which is believed. This information appears to be inherited from the verb. In other cases, such as John enjoyed (watching) the film, John began (reading) the book, or John finished (drinking) the beer, it is more plausible that the missing information is provided by the lexical specification of the; NP complements (cf: John enjoyed (drinking) the beer, John finished (reading) the book). Pustejovsky (1989, 1989b) and Pustejovsky &amp; Anick (1988) propose that the lexical representation of nouns is enriched to include a specification of processes typically associated with the objects they denote and that, in cases of coercion, this information is utilised. In their terms, this is the tdic role of the qualia structure of the noun.</Paragraph>
    <Paragraph position="5"> We see the inheritance of this information from the verb or complement as a default process which operates in the absence of more marked pragmatic information. For example, one would normally enjoy (watching) the play, but it would not be difficult to construct a discourse context in which someone (say lecturer or student) enjoyed</Paragraph>
    <Paragraph position="7"> (reading) the play, and so forth. So we propose that enjoy in this sense is a relation between an individual and ml event m-Kl that, by default, nouns such as fihn or play inherit 'watch' as a specitication of the typical event (process) in which they participate. The entry for enjoy will, also by default, state that in cases of coercion the specification of the process will be imherited from the nominal comt)lement. In cases where tim defanlt is overriddm,, by pragmatic information more specific instances of the entries tot enjoy and/or play are created in which the defaults are replaced by pragmatically appropriate ';pecitications. (The precise nature of the processes which trigger this or the retxicval of tile relevant intbrnmtion we take to be a lmrt of 'pragmatics' and not lexical semantics,) One tinal (mostly methodological) tx~int is that the approach we are advocating provides a slightly different viewtxfinl on the problem of lexical decomposition / represenlation. Early approaches to lexical semmltics within the generative hadition were criticised for the ,'ubitrariness of the representations produce*t. Following Dowry (1979), Pustojevsky (1989) and others, we suggest that one strategy for uncovering the optimal lexical representation, or level of 'decomposition', is to tx)sit representations which provide elegant accounts of the interaction of lexiegd semantics with grammatical realisation and with comtmsitional semantics. Pursuing this methodology, we have been led to a model of lexical semantic representation which suppo~ts a (somewhat emichexl) comtx)sitional account of (sentence) meaumg by enrichhlg lexieal ~epresentations of nouns and collapsing those for verbs with alternate grammatical realisations. In this framework, there arc. still many inferences which are not captured &amp;quot;tl. the level of lexiea\] organisatien, bu! we argtlt' that these inferences are. 'pragmatic' i~l the sense t\]la\[ they art: lh)f b'v-;ed on default processes operating within tim lexicon.</Paragraph>
    <Paragraph position="8"> Ttms, our position is OplX~S~{.d la lha:: of flobbs ct a1.(1987) who argue that there is no disti~ction between lexical semantic:~ and general knowledge, hi our approach, simple default 'lexicat' ifderence l~,rocedure.s do quite a lot of work. Of course, the way is always open to us to argue that rely inference which cannot be captur~xl by these procedures is 'non default'. New:Ilheless, in section 3 we argue thai this distinction is supt~;rted by natural data both in terms of the fruity of non-default cases and also the marked, info~mationally-rich nature of the contexts in which lexical defanlts ark overridden. Thus, our approach gives us a handle on which aslxacls el' k;xicN meaning sheukt be represented in the lexicon m~d therefore on the type of information we wmat te extract from our MRI)s.</Paragraph>
    <Paragraph position="10"/>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2. An Implementation
</SectionTitle>
    <Paragraph position="0"> It is possible to implement a system capable of' coercion and default specification using a unification-based formalism extended with 'orthogonal' default inheritance of (paradigmatic) lexical specifications. We also make use of minimal disjm~ctive specifications to allow for the range of grammatical alternation within one sense of a tmxlicate.</Paragraph>
    <Paragraph position="1"> Our prototype extend,; PATRolI (Shieber, 1986) with disjunction of atomic terms and uses the template mechanism to imtx)se a natural m~bsumption ordering on the lcxical taxonomy which defines the inheritance network. The taxonomy implicit in the fragment implemented so far is shown in Figure 1. q\]fis t~monorny is adequate to cover the metonymies discussed in this paper and others discussed in Pustejovsky (1989). (Numbers on concepts are relatexl to LDOCE sense numbers.) An entry for book is given in template from in Figure 2a. Its position in the network in Figure 1 defines the pattmn of inheritance for the qualia structme.</Paragraph>
    <Paragraph position="2">  The relic role fur book is thus inheritexl li'om file default role associated with 'Literature'. The entry will also inherit reformation from 'PhysObj' lint the orthogonality const~aint rules out conflicts with the attributes inlmrited from 'Literature'. In fact the template 'PhysObj' does not contain any information about the telie part of the qualia structure.</Paragraph>
    <Paragraph position="3"> The DAG for rite NP the book is &amp;quot;also shown in Figure 2a. This still denotes an object; when combined with a norrnal, non-amrcing verb the telic role makes no contribution to the semantic structure. However some grmmnar rules allow type-shifting; one allows NPs with an associated telic role to be type-shiftexl to be equivalent to untensed VPs mad to denote events, Figure 2b shows the NP after application of this rule.</Paragraph>
    <Paragraph position="4"> Once type-shifted, the logical formula associated with the book is the same as that associated with reading ttw book, except that the question mark indicates defeasibility and could be inteq)reted as 'possibly(P = read) &amp; P(e' j x)'.  In (la) we show the formula which can be read off the DAG in Figure 2b given straightforward assumptions about the semantic interpretation of the formalism (e.g. Moore, 1989). The lexical entry for enjoy specifies that its complement must denote an event which can be syntactically an NP or progressive VP and that, if the NP is type-shifted, the relic role supplies the understood predicate. The resulting formulae associated with the VP and S are shown in (lb,c).</Paragraph>
    <Paragraph position="6"> We follow Hobbs (1985), Alshawi et al. (1989), Moens et at. (1989) and others in using an event-based calculus for reasons of computational tractability, and also because distinctions amongst types of events are likely to be important in the characrerisation of the recovery of unc~rstood predicates in logical metonymies. In a fuller account it would be possible to constrain the type of event selected by a particular verb; for instance, enjoy might be constrained to unify by default with the telic role of a norm if this specified a process or culminating event. This would predict the relative oddity of examples such as John enjoys his house, in which we assume that the telic role is somettfing like 'living in' and that this specifies a state rather than process. It would also be possible to alter the aspect of qualia structure selected by a particular verb. An example like John regrets that book by default receives an interpretation in which 'writing' is selected to flesh out the metonyauy. In this case, we might specify that regret, in contrast to enjoy, selects the agentive path in the noun's qualia structure.</Paragraph>
    <Paragraph position="7"> Another area in which this approach to lexical semantics is suggestive relates to adjectival modification. It is well-known that adjectives such as good, fast, long, and so forth, have meanings which are hard to specify independently of some 'aspect' of the noun they modify.</Paragraph>
    <Paragraph position="8"> Pustejovsky (1989) suggests that in examples like fast car, fast typist, or fast waltz, fast should be treated as a modifier of the telic role associated with these nouns, so that: these examples can be paraphrased fast car to drive or fast waltz to dance. The adjective long appears to be (at least) ambiguous between a telic role modifier and a forrnal role modifier - a long book can either be a comment on shape, size or number of pages, or a comment on the length of time required for reading. In the event-based calculus we adopt we could associate the logical form in (2b) with the interpretation of (2a) where long is a telic role modifier.</Paragraph>
    <Paragraph position="9"> (2) a) John enjoyed the long book b) 3 e e' e&amp;quot; x y enjoy(e j e') &amp; ?read(e' j y) &amp; book(y) &amp; long(e&amp;quot;) &amp; ?read(e&amp;quot; x y) However, note that it would be inappropriate to automatically conflate the events e' and e&amp;quot; because this would predict that John's reading of the long book was necessarily a long event which, whilst plausible, is not entailed under this interpretation of long. In order to avoid this effect using unification-based techniques it is necessary to explicitly copy the structure that specifies the telic role. We suggested in section 1 that NPs, such as the fact, can denote propositions 'directly'. Similarly, we think that there is no metonymy involved in examples such as John enjoyed the experience /film-making and so forth. In these cases, we claim that the NPs in question denote events 'directly'. Thus, we are lead to an 'ontologically promiscuous' semantics (Hobbs, 1985). However, recent developments in model-theoretic semantics which treat properties as basic entities (e.g. Chierchia &amp; Turner, 1988) support this position. Indeed the interpretation of event-denoting NPs in complement position with enjoy strongly suggests that these NPs must be analysed as denoting propositional functions since their 'missing argument' must be associated with the subject of enjoy. For instance, John likes marriage can mean that John likes the institution but John enjoys marriage can only mean that he enjoys being in the state of marriage (to someone).</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3. Data concerning Logical Metonymles
</SectionTitle>
    <Paragraph position="0"> The previous sections have demonstrated the nature of the phenomenon of logical metonymy and have outlined a computationally-tractable unification-based trealment. A crucial aspect of this treatment is that, with the predicates we have considered, the missing information is supplied, by defatflt, by the qualia structure of the head noun in the type-shifted complement. In order to demonslrate the presence of logical metonymies in naturally-occurring text and to evaluate the plausibility of our default approach, we examined data drawn from the Lancaster-Oslo/Bergen (LOB) corpus containing predicates capable, in principle, of coercing the type of their complements.</Paragraph>
    <Paragraph position="1"> A set of type-coercing predicates similar to enjoy was obtained by extracting verbs coded to take both NP and progressive or infinitive VP complements in LDOCE (see Boguraev &amp; Briscoe, 1989b for an account of these codes and the extraction techniques). Further manual editing of this list led to 24 predicates which we felt were capable of exhibiting logical metonymies parallel to that of enjoy. To date, we have analysed all the data obtainable from the LOB corpus for seven of these predicates. The results of this analysis are summarised in Figure 3. (Numbers after predicates refer to LDOCE sense numbers.)  the number of times each predicate occurred with this tYtVe of complement. (A stroke in fltese eolunms indicates that this complement type w(mld be ungrammatical with a particuh~r predicate.) The remaining cohunns give further information about the NP complements. Ev(cnt) indicates the ntlLmber of times that the NP complentent was judgexl to denote an event (or hi a few cases a proposition) directly: Met(onymic) indicates the ntunbar of times we judged that coercion had occmred. And Prag(matic) indicates the number of times that we judged the understood predicate was not recovered via the head noun's qualia structure in the metonymic cases. In some cases, the number of NP complenmnts is greater than the stun of Event and Metonyntic because we felt unable to classify some exmnples. These examples were either (senti) idiomatic, such as miss the boat, or involved NPs whose status was unclear because of modification of the head norm, such as enjoy the warm evening.</Paragraph>
    <Paragraph position="2"> The first thing to note about Figure 3 is rite comparatively high numbers of metonymic exmnples relative to the complete sets. It is inslructive that the apparently more complex metonymic complement pattern is selected quite frequently despite the availability, with ~dl ttmse 1)redicates, of an explicit VP complement pattern. For instancx~, enjoy and moq~hological vmimlts occurs 65 times in the relevant sense mid coerces its NP complement in 25 of these cases. The second and crucial ohse~v~,tion, from the perspective of our default theory of the recovery of the understood predicate in the metonymic eases, is that the numbers in the 15ragmatic column m'e relatively low by comp~dson with the total nuntber of metonynfic exmnples.</Paragraph>
    <Paragraph position="3"> Given the defimlt theory, we would expect most rnetonymie examples to be resolvable via the head noun's qualia structure and there to be relatively few 'pragmatic' examples involving less constrairted lind ntore complex inferences, and, in fact, these cases represent aixmt 17% of the metonymic examples and about 4% of the total set of exanaples considered. Further examination of these 17 cases revealed that, in most, the hnmediate context was informationally-rich arid therefore marked enough for the appropriate pragmatic inference to go through. For example, compare the a) examples with the b) examples in (3).</Paragraph>
    <Paragraph position="4"> (3) a) Willie enjoyed the hot sweet te~h standing on the deck in the cool of the night b) Site can lie back and enjoy her baby tmtil the midwife, knowing the afterbirth is ready to pop out, a) Loddon paid Iris own account, finished lfis cigarette, and got up.</Paragraph>
    <Paragraph position="5"> b) The Ix)ok was never finished, for his illness ~md death intervened while he was in the course of wriling it.</Paragraph>
    <Paragraph position="6"> a) If you prefer a Burgundy try a 1955 Charmes Chambertin costing round 1 txmnd.</Paragraph>
    <Paragraph position="7"> b) Then again so rnany t~ople much prefer the sea or river to the 1)aths. Having learned to swim in the sea ...</Paragraph>
    <Paragraph position="8"> in each of the a) examples we think that the tm(terstood predica|c is supplied via the qualia structure of the head noun in the NP complement, ht the b) cases, it seems implausible that the telic role of babies is to be cuddh;d, or that seas or rivers are (mainly) for swinuning in. &amp;quot;\['he agentiw.&amp;quot; role of book will specify the predicate 'wriuz', st) we could treat finish as selecting this role by defmdt m~d this would, in fact, deal with four of the 'pragmatic' cases, but othexs would become 'pragmatic' since in our implementation only one unification path hato the qualia structure c~m be selected by default. However, hi each of these examples the context shown provides enough information to infer the relevant predicate, it is not ~flways tim case that the (remah~der of the) context provides the relevmit information or intuitively seems so 'rich' in the default cases.</Paragraph>
    <Paragraph position="9"> /MLother way in which we can evaluate the default theory is by considering the status of the predicates which are supplied explicitly when a VP complement is selected. We might expect VPs to be selected precisely in those situations when defimlts based on qualia structtue would lead to the wrong intewretations. We tested this idea by exmnining the VP complements of start. In many cases, the prtxticates were intransitive, dilransitive, and so lbrth, so that rite hypothesis did not apply, tlowever, in the srraiOtfforwardly trmtsitive cases 21 exmnptes exhibited clear non-default pre(ficates, such as started to open the bottle, started to play a Waltz, or started flirting with the first pretty girl that you met, whilst only 4 cases arguably inwglv~l defmflt predicates recoverable from the head ~otm's qualia structure -- start making a fuss, started to fire disitress rockets, started pulling the commtmication cord, artd started making bubbling twises.</Paragraph>
    <Paragraph position="10"> This analysis is hardly conclusive, however it doe';, we think, demonstrate that logic~d metonymies occur quite regularly with certain predicates in natural text. We have also provided some evklence that default inference based on lexical organisation (in this case the qnalia structure of nouns) would succeed in a large number of cases.</Paragraph>
    <Paragraph position="11"> Furtheimore, there seems to be some support in this data for the claim that contexts in which 'pragmatic' recovery of the understood predicate occurs are quite infbrmationally-rich mid would therefore constrain m~ otherwise rather unconstrained process. Finally, we have shown that, in the case of start there is evidence that VP complenrentation is chosen when default recovery of understood predicates on the basis of qualia structure would lead to the wrong interpretation.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4. Acquiring Lexlcal Semantic Information
</SectionTitle>
    <Paragraph position="0"> In this section, we describe three exploratory studies aimed at tile (semi-)automatic acquisition of qualia structure, in particular telic roles, fTom lvlRDs. The first involves exploiting subject and box codes in the LDOCE MRD (see papers in Boguraev &amp; Briscoe, 1989 fbr a full description), while the second is based on an analysis of the LDOCE definitions. These teclmiques are aimed at allowing qualia structure to be inherited appropriately; the flfird attempts to determine the predicates associated with a word by analysis of dictionary definitions and, to some extent, more general corpus material.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 45
</SectionTitle>
    <Paragraph position="0"> The machhae-readable version of LDOCE contains some residual 'database-like' features which do not appear in the printed dictionary. These include a taxonomy of many words in terms of 'subject matter'. This taxonomy defines a 'fat' hierarchy of, at most, two levels and many relationships are left implicit; for instance, 'sports' is a main extegory with subdivisions such as 'archery' but 'football' is a main category with subdivisions such as 'rugby'. Nevertheless, this taxonomy can be used to identify 'lexical conceptual paradigms' (Pustejovsky &amp; Anick, 1988); for example, there is a class 'beverages' (147 word senses), a class 'motion pictures' (113 word senses), and a class 'literature' (377 word senses). These words c.ould straightforwardly be associated with the 'deeper' inheritance network given in Figure 1 with default telic (and possibly other) roles, such as 'drinking', 'watching' and 'reading' associated appropriately. There are a few problems though, for instance the category 'beverages' includes publican, and 'motion pictures' includes usherette. It is possible to exclude these examples from the target network by making use of box codes which, amongst other things, associate semantic features with nouns, because the exceptions are coded 'animate' and 'hmnan'. Nevertheless, this approach is limited because the LDOCE semantic taxonomy will undoubtedly not contain all the classifications which eventually will prove desirable and there will be errors of omission in its construction. In addition, we are ufilising an idiosyncratic feature of the LDOCE MRD, wlfilst we would like our extraction techniques to be generally applicable.</Paragraph>
    <Paragraph position="1"> An improvement to this approach is to utilise taxonomies constructed from the dictionary definitions. For example we have built a taxonomy of substances by extracting the genus senses of LDOCE definitions in which 145 word senses such as Burgundy appear directly or indirectly under the main nominal sense of drink. We are currently investigating an approach whereby lexical entries inherit some of their structure from higher nodes in the taxonomy. Qualia structure could thus be inherited from word senses rather than abstract templates; for example Burgundy would inherit its telic role from the noun drink.</Paragraph>
    <Paragraph position="2"> If abstract templates were still needed they could be inserted into the inheritance hierarchy at the appropriate points.</Paragraph>
    <Paragraph position="3"> The approaches above only specify how the qualia structure is inherited, rather than how it is initially det~'mined. In recent work, the IBM lexical systems group have used their lexical database system (e.g. Neff &amp; Boguraev, 1989) with a number of MRDs to generate lists of pre~licates which are applied to books by searching through definition fields for the occurrence of book in a position denoting 'typical object' of the headword. For instance, LDOCE defines sag with '(of a book, performance, etc.) to become uninteresting during part of the length'. Using these techniques with three dictionaries resulted in the following list of verbs: abridge, abstract, annotate, appreciate, autograph, ban, bang about, borrow, bring out, burlesque, bowdlerize, call in, castigate, castrate, catalogue, censor, chuck away, churn out, classify, collate, commission, compile, consult, cross-index, dramatize, entitle, excoriate, expurgate, footnote, page, pirate. It is obvious that this technique yields specific, often rare, predications with typical objects. Whilst qualia structure is likely to involve typical predications with specific (classes of) nouns.</Paragraph>
    <Paragraph position="4"> In Order to automatically obtain typical (frequent) predications of book, four corpora were searched LDOCE example sentences, the Brown corpus, 1.2 million words of Readers' Digest, and 26 million words of tapes from the American Publishing House for the Blind.</Paragraph>
    <Paragraph position="5"> Analysing those citations in which book occurs as direct object revealed that read and write are the two most common predicates across the four corpora, although ~l~re are considerable differences within each corpus '~see Boguraev et al., 1989 for details). This approach could and should be extended in several ways, for instance by dealing with semantically related nouns such as novel, and, of course, by attempting a similar analysis for many more nouns. Nevertheless, these preliminary results do suggest that a noun's qualia structure should be recoverable from MRDs and corpora in a semi-automatic way.</Paragraph>
  </Section>
class="xml-element"></Paper>