File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/94/c94-2112_metho.xml
Size: 20,544 bytes
Last Modified: 2025-10-06 14:13:41
<?xml version="1.0" standalone="yes"?> <Paper uid="C94-2112"> <Title>On the Proper Role of Coercion in Semantic Typing</Title> <Section position="4" start_page="0" end_page="708" type="metho"> <SectionTitle> 2 Polymorphic Languages and </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="708" type="sub_section"> <SectionTitle> Semantic Expressiveness </SectionTitle> <Paragraph position="0"> We will a, ssulne s(3me geueraJ familiarity with the framework of generative lexicon theory, as outlined in \[16, 18, 1\]. ~'e feel it is important, however, to clarify the motivating principles and genera.l methodology behind SllCh work, since these points seem to be overlooked or misunderstood by some authors (\[7\]).</Paragraph> <Paragraph position="1"> In order to help characterize the generative power of natura.l languages in terms of semantic expressiveness, it is natural to think in terms of senla.ntic systenls with increasing functional power, l:urthel'more, a nat u ra.I wa.y of Cal)tu ring this might be in terms of the type system which the gra.mmar refers to for its interl)retation. It has been argued elsewhere (\[19, 20\]), that there are reasons for describing how semantic systems fall on a hierarchy of increasing expressive power.</Paragraph> <Paragraph position="2"> It seems clear at this point that the current enumera.tiw ~. techniques for lexical description are too impoverished to adequately describe the richness of semantic da.ta, much less to explain either how word senses relate to one a.|tother or the creative use of wo,'ds i. ,,over ,:o.texts llrieily, a. genera.tiw', lexicon can be characterized as a syste.nl involving at least the tbllowing four levels of re.presentations: Argument Strut 1;nre, Event Structul'e, Qnalia. Structnre, and I,exical Inheritance Structure. A set of generative devices connects these \[bur lew~qs, providing lbr the compositional interpreta.tioll of words in context.</Paragraph> <Paragraph position="3"> The exact nature of these devices will determine the polym,m'phic expressiveness of t, he semanLics in fa,irly delinite ways. The best st.u(lied illustration of this is the phenomet,on of ::+\]p++ coc?'cio?+, but it, is by no mea.ns the only one.</Paragraph> <Paragraph position="4"> &quot; (.oer(lOlt 2.1 LinguisLic Evidence for &quot; As ment, ionell in \[18\], the t>henotnunon of multi l)le subcaLe+goriza, tion has tnotiva,t(+(l mu(:h of lhe type cha,\[lging literature. The alq)roa,ch taken itt getmrative lexicon l, heory build.,, ou tire ideas de veloped iu \[13\] and \[9\], while at:t,m~l/ting to derive the synta,ctic expressior~ ,:)f a verb's l+omplemeut: on the basis of a. deell senl,,inl.i(: type ;-issigl,tllellt, together with synta.cLic constraints, l:or examl)le, i0+ the we, ll-stludied case of aspectual verb conlplementation in (\]) a.nd ('2)below, the verbs bc:\]i++ a,n(\] (:O~ll, llto'lL('+'r (;a, rl'y a (lee I> l;ype s(qecl:in~ for at1 event in (:oJnplement F.osi/,iol~.</Paragraph> <Paragraph position="5"> (I) a,. John h<'ga, to read the hook. (vP\[-t INI,'\]) h. ,lohI~ I)egau reading lh<' ho<>l,:. ((',1') c. Johu llega, n the huol.:. (NI') (:2) a.. Jea.t+ a cotuulen(:(' 5. tire l(' tivre.</Paragraph> <Paragraph position="6"> I1. Jean a (:oumten<:(, le livre, This <l(~ep I~yl)e ix able 1:o liJ'<>je(:l: to one o\[ 1,hree possible surl'ace forms, <h,pen<liug ou whi(:l~ coer don ,'ule a.l>lllies (\[IS\]), There is, ho,vewu, only one semantic t:yl m being seie<:te(I rot', aud tile <:hlstering of the part.i<:ula.r synl.:-icl.i(: Ibrms a p-Ilea, ring a,s surl'ace <:,.)tuplem(,iH, I,Vl>(,s iu (1) ;1re sysl, ema,ticatly I)rojected t~y virtue of this sema, tl tic t, ype. That ix, auy verb, like b+:li++, s,.,lectiug for a,u &quot;u nsa,L m'a,ted (wet, t', will parad igmatica, lly allow for the expression or the three gramma,ti.</Paragraph> <Paragraph position="7"> ca\] forms showu a hove, a.ssuming surface syntactic consl, ra,inl:s a, re sa,l, isfie(l. For this reason, ~:he stru<:ttu'ing of this kiud or l(nowle<l~/,, where t lli~ event type ha,s syr~ta,,::l, ic exl)ression as ,any one o1' the surface types in (1), is called a hzi,:',l ,'o~+<:,7~t~talparadi:\]m (lop). In this view, t, he NP, a boo\],:, is ('oer<:e<l to Lhe :.qipr<)l)rial;e t,yF, e re(luire<l hy its governing verh. Wha,t m:C/l.:es coercion l),Jssil)le in this <:a,se is the awfilability ,.ff lhe required type, given as pa,rt of the N l>'s qualia sh'~whu'c, in(If <:a,ting, for exa, ml>le, I, ha t t, he 'i'\],;I,\]<: r(>le for book iS tile activity (1\[' rea.dittp,, while Lhe AC;I')NTIVI') role ix IAte al:L (Jr wriLiug. Tim result of apply ing this coercion (:.l)erat,:)r t<> an N P ix (,ff('/:l.i,:(%' Lo crea, te ~n cxlc~,,,'#m, o1&quot; 1,he N l' mt'a.tfing, calMl a 7uchngumic rccoTz.st?&quot;uc.tio?~+. In the ('a.se o\[ Lhe NIl, (z boo/G for exa41H)\[(: + Lhe o\])(:.ra.gor prc.ditces illisagtlt'atc, d 0vot/t (\[(HIo|z4,LiOllS.</Paragraph> <Paragraph position="8"> There .h.re s(!v(:ra\] phe.ltOlnetta dis('ttssed iJt (',o ~ dard and Jayez IT\]+ which they c+lahn illustrate thaL coer(:iOll is noL a viable iut;erpregive sgraLeg:y for linguistic selna.ntics. All, hough none of ghese al)pa.rent, cou tlt, erexam l)h,.~ is in fact a l)rohlem for (;en(,ra.tive I,exicon Theory, it. is itnportant to dis russ each I)He\[ly to show why they are false \[)rob \[(!tll5, &quot;~\:0 \ViH COllCelltra.L(', howov(':r., oil \[,ho s.L,\](!ctiona\[ prol)erLies of aspect:ua\] verbs stlch as co?n?uc+z(++r and bc(\]i?+, in order t,o show very ('le+u'ly thaL sense enutl~era.tive a.l)proa.ches such a,s (++o+ dard a ud Ja.yez's are tnimsing Lhe poinL of li+}guis tic and computal.ional generaliza.tions, as regards l,o how t, lle lexicon (:ottLril)ttl, es t,o t, lie (:otnpos\]+ ti(mal setnautics.</Paragraph> <Paragraph position="9"> Th(, first, a, ppar(,ut, couttt, erexa.ntt)les , (liscuss,.+d in (',odat'd 'aud Jayez (\[7\])+ t,<) the ~e.eral a,pplica, tion or 1,ype chauging oF, era,tions show tha, t, ('om, m+,~c+r does not utfiwu'sa.lly allow NI' (:ore.</Paragraph> <Paragraph position="10"> i)l,:,ments with a <:oerced inl:erpreta,tiou, l&quot;or ex+ aml)le, the Nl's in (3) below do uog have t, he expected eveul; readings that; one would predict, were, thore no i:o\]tsLra.inl;s ,:)tl tahe apl)lica,tion o1' l:yl)e (:u(,r(:i(:,tl o\[)era.tiolts.</Paragraph> <Paragraph position="11"> (3) a. *Jean a. co,t,,,,e,,c(+ t, ne sy,npho,fie (\[7\]). &quot;.John I>ega.n \[to listen to,\] a symphotiy&quot;.</Paragraph> <Paragraph position="12"> 1). :~\[~\[;11'i(! D+ COlIIIII(!IIC(&quot; \]~ltlLOl'Ottl,(L &quot;Mary began \[t.o (lrive (m\] t.he highway.&quot; c. *.l<)hn begau the <lict, iouary (\[16\]).</Paragraph> <Paragraph position="13"> &quot;,John bega.n fro c<)nstflg/reference\] t, he dicLionary.&quot; null But, as already l)oint, ed out iu \[.9\], the a,ccel)ta,tfility of coerciol\[ wit\[i aspec, tua\] l)re(li(:~ges such as comn~(~?wcr and bcqh~ is conditioned 1)y Lhe te\]i(:\]ty of Llle (went t, aken as it;s coln\])\[em(mL. Briefly, these verbs s,,qe(:L \[or an event of the sort 'L'RAN.~\]TION, rtlli\[l~ OUl. Lhe coer(:ed illt(.'rl:,reL:.l.Liotls o\[' li.,+:r~?~ Z<., for (3a.), :h'ivc o7~+ \['or (3b), and co~.,,&quot;Itlt For (?,c), wlfich are a.Jl I>RO(H,;BB ovenLq. \]:'urtJlor tnore, cotlstrainL~ due Lo &quot;l)oundedttess&quot; o\[' t, he pre(li(:a.Le ((,la) vs. (41))) are enL\]rely consist;('nl: \viLli condit.iOtlS oil (:Oel'CiOll ill (;(~'l\](',t';;tlSjv(+ \],ex$ co,, Theory ((:f. \[L(i, 17\]).</Paragraph> <Paragraph position="14"> (4) a+ Jean a commencd \[e t'romage. / le livre.</Paragraph> <Paragraph position="15"> &quot;John I)(!gan the cheese (c, atiwj) / the book (reading).</Paragraph> <Paragraph position="16"> b. *Jean a commenc(~ du frolnage / des livres.</Paragraph> <Paragraph position="17"> &quot;John began cheese (catinq) / 1)ooks (reading). null Namely, the homomorphic rela.tion between the NP type (mass vs. count) gives rise to i)rocess and transition interpretations of event structures (corresponding roughly to the amorl)hous and bounded readings resi)ectively, of (k)(lar(\[ and aayez's analysis).</Paragraph> <Paragraph position="18"> If it is truly an expla.na.tory and productive operation, coercion should be trot just a property of object phr~tses, but affecf the semantic inter-pretation of subjects and other positions a,s well (cf. \[17\]). For exan, ple, the interpretation o|' psychological predicates such as ill (5) involves a inetonymic reconstruction of the subject as a.n e.xperiencing event.</Paragraph> <Paragraph position="19"> (5) a. Books bore me.</Paragraph> <Paragraph position="20"> b. The movie Mghtened Mary.</Paragraph> <Paragraph position="21"> c. Ma.ry's ihce / her chatter / listening to l'V\[ary bores nle (cf. \[7\]).</Paragraph> <Paragraph position="22"> Contrary to Godard and Ja.yez's claim, all of the above examples indicate very clearly a sub-ject event reaming; i.e., ~va(ling books, watching the movie, seeinq Mary's face, and listening to her chatter, as argued in \[J6\]. With examples such as *The book bcga'n last week, however, co ere|on is not possible for ra.ther trivial reasons; nalnely, as a violation o1' control. It is, by tile way, not surprising to lind asymmetries between argument positions. Anaphora., control, and extraction from subject position all behave differently from argument positions within VP. The point is that linguistic evidence supl)orts an underlying semantic type, directly explaining what the connection between the subject and object of the experiencing relation is. in \[17}, the un(lerlying semantics of psychological predicates such as bore, anger, and frighten is a causative strue lure where the surfa.ee subject is the logical object of an experiencing relation. For example, the event structure for the verb an.qcr has tire tbllowing form:</Paragraph> <Paragraph position="24"> The qualia structure projected by the NP cow tributes relational information as to just wha.t manner of experiencing is inw)lved. Short of general world knowledge, how are we to infer the 1)t~r titular ma.nmu' in which Mary became bored in (4b)? By knowing what a. movie is, we know how to use it and experience it; this is the defining role of tire quaJia structure. And yet, to claim that the qualia are a useful representation (which C,odard a.nd Jayez admit), withont exploiting them through type reconstruction operations (e.g. co ercion), is to fail to see the logical relations be~ tWeell lexical senses a.nd derived senses in tile language. It is as though we were to i)ermit traces ill our granuna.tical formalism without having a.</Paragraph> <Paragraph position="25"> statement of binding, or otherwise knowing what to do with them.</Paragraph> <Paragraph position="26"> The third argu|nent aga.inst (:oercive opera.tions involves exa.mples such as a long novel and a bright bulb. These. are to be contrasted with a red book and a,i opaque bulb. As pointed out in \[21\], the adjectives here modify a distinguished event predicate (i.e. a quale) associated with tile head, rcad for book, a~td ilh, min.atc tbr bulb. Godard a.nd Jayez seenl to think that because the NP can apl)ea.r in an environment typed for an individual, such as (6)below, (fi) .lean a. achet6 un h)ng roman.</Paragraph> <Paragraph position="27"> &quot;.John hought a long novel.&quot; tha.t this is a. counterexample to type coercion.</Paragraph> <Paragraph position="28"> But this surely misinterprets what role the adjective is playing in tile semantics. As already argued very explicitly in \[22\], the modification by an adjective such as long, rclpide (J?tsl), or brillant (bright), is a submodification on the al)propriate qualia of the head.</Paragraph> <Paragraph position="29"> (7) Ax\[rom, an(:c). ,.A \['/'~ic(x)- ,\,l,,\,,r\[ lo,~v(# r) \]\]\] lire(e T, w, x) A The resulting coinlmsitiona\] structure is still the type of tile whole NP, and has no effect whatsoever on selection by an outside governor such as acheter (buy) as in (6).</Paragraph> <Paragraph position="30"> &quot;File \[ina.I signili(-ant argument Go(lard a.nd Jayez l)resent against coercion operations inw)lves the apparent lexically idiosyncrati<: na.ture of coercion. Why should commencer an<t finir allow eoercion while cesser and arr~.ter do not? There is no space to detail the distinction here, but it is ai)parent that this is (hte to a selnanti(; tyl)e distinction between these classes of pl'edi (;ares.</Paragraph> <Paragraph position="31"> In what follows, we demonstrate how t, he al)parent violations of the coercive behavior of begin-predicates actually reveal a much deeper se mantle distinction t)etween two logically relate, d senses of the verb, in all the complement fo|'ms they take, and not ,lust NI ) coml)lement, cases.</Paragraph> <Paragraph position="32"> This can be applied mutatis mutandis to eoml)Zg~l, CCl',</Paragraph> </Section> </Section> <Section position="5" start_page="708" end_page="710" type="metho"> <SectionTitle> 3 The Semantics of begin </SectionTitle> <Paragraph position="0"> As argued in Section 2 al>ove, the well-formedt|ess of object coml)lement coe+rcion with ast)ectual predicates such as begin is ColMitioned by the event sort of the (lualia a.ssociate(1 with the NI ) itself. Thus, only Nl)s having associated tra.nsi+ tion events will allow coercion a,n(\[ control. This is not to sly, howew~r, that bcgi?z selects only for transition events. There are, of course, perfectly grammatical examples of prt)cess COm l)lelnents, as shown in (8) below: (8) a. The snow began to \['all at mi(llfight.</Paragraph> <Paragraph position="1"> b. John 1)egan to feel ill.</Paragraph> <Paragraph position="2"> c. The W~/l\] ' bega.n to t'each ilttO Bestride. These examples illustrate the use of begin as a raising verb. We will follow Perhnutter \[14\], in distinguishing between two senses of the. verb begin, distiuguishal)le not I)y the selectional properties given in Codard and .layex. but, rather, con+ forming to the distinction that \[1.t\] ula(le; namely, ~'~S eit\]:ter a Raisi'ng or a (~'o?~lrol verb.</Paragraph> <Paragraph position="3"> The analysis is as folk)ws. There are in(leed two grammatical expressions of the verb &:gin, as Raising and Sul)ject-(,'outrol forms: As a control verb, the event sort specified as tim c(maplement iS a TR+ANSI3'ION. As a Raising verb, however, the event may be any sort. This tbllows the' typing assignments below: Controh ((' .... (r) _ (( _ (r) Raising: (c deg / t 'r) The examples above aud in (9a) and (9b) I>elow illustrate the raising i||terprt~tation (~\[' bcgi~: (9) a. The a.cid began to corrode the marble.</Paragraph> <Paragraph position="4"> b. It. began ~o rain.</Paragraph> <Paragraph position="5"> We will assume that ra.ising is accomplished by function composition, in the manner of \[8\]. The manner in which Raising is treated a.s function composition (FC) is as follows: begin is ~:* -~ ~'*', to corrode the marble in the exam.ple above is e -+ c t'. Then, b'C(be(jin , VP) -a?,\[b,:vi,,,'( ,,o,,,,oa,:( ?-,, O...-,..,.q,l~)\]. As pointed out irt \[?\], VP ellipsis can be used as a diagnostic for determining whether a conq)lement is l)art of a raising or control construction. Some l)redi(:ates permit 1)otl\] a control and non c()ntrol reading, such as (10)below, where John may I>e intentionaJly (lietitlg or he may be ill. (10) .\]ohn began to lose. weight.</Paragraph> <Paragraph position="6"> Notice however, that in English the sentence in (1 I) has only t.he intentional inclloative reading, aim not the raising version.</Paragraph> <Paragraph position="7"> (11) JohI |began tolose weight, and Mary began tot).</Paragraph> <Paragraph position="8"> What this illdicat(,s is that there are indeed two coJlstru(:tioJts at play here, as teased apart I)y cert, ain diaguostics. I&quot;urther evidence comes ft'o|n im peraLive struct u res (12) an(l force-corn plement (:onstructions, which require tire control sense of the verl).</Paragraph> <Paragraph position="9"> (12) a. *llegit! lookitlg for a.iob, you lousy bum! b. Start Lot)king \[br a job, you lousy bum! These data iIMicate that, begi?z, in the control interl)retation , strongly preli~rs a relic (transition) evelLt cOllll)\[enlelll;.</Paragraph> <Paragraph position="10"> We have argued that there are, two senses of the verb bc(li?~, cor|'espondit|g to raising a.ud control predicates. These senses, howe, ver, are not arbitrary t;ypes but are logically related t,o one an other in the same way that tile diffe|'ent senses of una.ccusa.tive/causative verbs, such as break, and sink are. related. In \[23\] it is shown titat verbs such as .~i~k and a./.fonda'rc are logically polysemous in predictabh~ ways, and don't need to be assigned multiple texical entries. The same geu eralization hohls for verl) such as begil~: begirt, is the lexical version of a,n u naccusative marker, 1)ut for propositions rather than ti)r entities.</Paragraph> <Paragraph position="11"> l,, \[~2\] +,,.1 \[~a\]. +, ~e,.erat ,.echa,,is,,, is ,It,fined which makes the appropriate tyl)e a.wdlM)le for a ('oe|'cion operation. As disc.ussed in \[16\], turning the value of a particular quale for a.u NP. The combined set of quaiia provide a set of type aliases for tile expression contaiuing them (of. \[18\]). One particular mechanism, type pumping, has been explored as a means to generating the alias set (\[20\]).</Paragraph> <Paragraph position="12"> l,et G be the typing judgements with respect to a grammar. Then, by convention, C I- ~, : r represents a type assignment of r to the expression C/~.1 Thus for example, the type a.wdlable to an expression o' with quale Qi of type r, can be see,) as the following type inference: 2</Paragraph> <Paragraph position="14"> This says that, given an expression a' of type (r, there is a coercion l)ossible between (r and r, which changes the type of a' in this composition, from (* to r. We will illustra.te tile further a.1)plication of this coercion ol)eration below, as used in the begin examples. \[n (13a), we see how the a.spectual verb commencer selects tile COm l>lenmnt VP, and how in (:lab), an NP is coerced into an event interpretation.</Paragraph> <Paragraph position="15"> (13) a. Marie a commenc6 5 lire le livre. (vp) b. Marie a commencd le liw'e. (NI') Following \[2()\], we can view tile l)a.sic COml)osition of the sentence in (13a.) as type inference in For the deriva.tion of (131)), coercion applies to the corn plement N \] ), resulting in the a.l)propriate type selected by the verb, as illustra.ted below ill Figure 2) In the case of b(;gin with NP co)nplements such as the symphony or lhe molorway, the <:oercion is not possible, given the type misma.tch in the intended qualia relation (i.e. li.stcning and driving are PROC\],~SS evelH;S). Notice, however, since the AGI~;NTIV\],', for each has all evellt of sort TRANSI-TION, these ;q.Fe possible coercive interpretations; i.e. perform the symphony, or build the highway. Notice that one might expect there to l)e raising constructious involving coerced NP com\])lements. But these do not exist:, as the ungrammaticality o\[' *John &tan his ~ap (non-control reading) illustra.tes. This ix due to the fact that coercion is governed by the type of the controller, in this case C' type c~ ~ ('c. This coercion will be successful if suc}l a type exists in the alia.s set of the complemellt. Since function composition is an ol)eratiou at. the level o\[' the VP, there is no point iu the deriva.tiou such that the. api)ropria.te type is availal)le \['or the rule to apply.</Paragraph> <Paragraph position="16"> As a filial ol)servation, it should be ol)vious now why verbs such as enjoy allow a nluch I)roader range of complement coercious (of. \[16\] for (h> ta.ils). They are typed for taking an event of any sort, thereby allowing the I)ROC\[';SS evel/ts Of' the &quot;\['l.;IAC roles ill enjoy lhc symphony / lhe movie. :~We ignore for now the t;yl)(', disl.inct;ion between individuals, e, aud generalized quantifiers, <<e,t>,t>. In I;he full version o\[' t.hc l)al)er, we show the type shift taking this clist.iu('l.ion into ~ICCO\[IIIL.</Paragraph> </Section> class="xml-element"></Paper>