<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2110">
  <Title>Virtual Polysemy</Title>
  <Section position="4" start_page="696" end_page="698" type="metho">
    <SectionTitle>
2 Lexlcal Polymorphism and q'ype
Resolution
</SectionTitle>
    <Paragraph position="0"> Our lloints of dep~Lrture are. (i) the polymorllhie approach to lcxical specilication of I'us/.ej(wsky (1991, 1993) aud (ii) the Attribute I,ogic Engine (AI,I;:) for realism dc.veloped by Carpenter (\[9,()2;t, t992b).</Paragraph>
    <Paragraph position="1"> Following Pustq\]ovsky, we adopt an integrated muL tilayered representatiou ol7 word meaning which incof porates salient aspeets of world knowh.'dge and *viler(; different use,~ of the same word are conllatcd into a sillgle mela-e'nlry. For example, a verb eutry is ~msigned ~t lexical type which provi(les a sl&gt;ecifie~ttlon of both argument and event stru&lt;'ture in&lt;:luding them~tic and collo&lt;-atioual (e.g. qualia) prol&gt;&lt;~rties &lt;&gt;f its parti&lt;:ipants ~tnd can be extended to achieve contextual congruity (see below). In contrast with l'ustejovsky, however, we do not Alse coercinu as a lrlaiil generative device to enl(.)rce seuse extcnsk)ns. True coer(,.ion hwolvcs \[,y\[)e shifting which is operationally equivalent to a lexicM rule (Pustejovsky, 1993). Consequently, the gener;t-Lion of sense exten,dons by coer(:km is ultimately of little avail in redueing lexic;d amlfiguil.y, a.'~ w~s noted earlier R)r lexical rules.</Paragraph>
    <Paragraph position="2"> \]lather than using coercion, wc encode lexi(:al polymorphism by type underspecilieation and generate sense extensions using contextual iulbrmation to ground lexical items. Wc provide such a simeillca.</Paragraph>
    <Paragraph position="3"> lion of lexical structure within (;arpenter's ALE using a tlPSGdike grammar I'()rmalisnl (Pollard &amp; Sag, 1992). This grammar formalism integrates a neo.</Paragraph>
    <Paragraph position="4"> l)avidsonian approach to verb semanties (Parsons, 1990) where thematic roles are delined as prototyl)ical notions (Dowry, 1991), sec Saniilil)l)O (1993). l,cxical types are ~m'anged into an inheritance hierarchy with l)olymorl)hic types ~s intermediate nodes; caeh type can be ~ussociated with cotmtraiqts expressed in terms of attributc--wdue pairs. For exanq)le, the lexical type of SylIS(',III for all intl:allsiLivc verb sileh }l,~{ swint is de flued so as to subsume i.ll(', types iv_,ndir~synse.nr and iv_obl_dlr_synse.m which characterize the two uses of the verb exenq)lilied in (ld). This is shown in the type lattice fragment in Fig \[ where . upl)er-('.~me characters are used lor attrilml;es and bold lower-.ctuse for tyl)es (many details are omitted tor e~u. ~, of exposition) (lyn_L've is a sort for non-stative eventualities (i.e.</Paragraph>
    <Paragraph position="5"> it subsumes processes and relic events) -pred is either a lexical or logical predicate (l(:x_pred, e.g. swim; log_t)rcd , e.g. aud) loc_chng is a thematic sort which characterizes participants undergoing change of location (lir_t)re. 1) is a sort for prepositions which express a directed path (e.g. to, acTvss).</Paragraph>
    <Paragraph position="6"> Because swim in the lexicon is assigned tile underspeciticd type iv_reMit orAv.x)bl_dlr~synsem, it can potentially combine with a complement and the subject arguments, or the subject only. In the tirst case, the complement list would he non-empty with its head instautiat, ing a pp_syns(~m (prepositional phrase). The value for the t)ath SYN:LOC:COMPS would thus resolve to the type 1)p_compdlst which as shown in (I) is the singleton list containing a pp-syIlsem.</Paragraph>
    <Paragraph position="7"> This is simply because e_or4qu:ompllst is defined as having subtypes (;.Jlst the empty list -- and pp_coml)_llst as shown in (5).</Paragraph>
    <Paragraph position="8"> Ill a typed feature structure formalism with gcneralizcd reeursive type. resolution (Pollard &amp; Sag, 1992:ch. 1; Carpenter, 1992a:ch. 15), the grounding of e._or._pf)_COml).Jlst to l)l)_COml)Aist wouhl sulIicc to solve iv_undir_orAv_ol)l_dir_synsem to iv_obl_dir_synsean, lnstantiation for tim head of tile compAist during l)~trsing would then be sufficient to detC/:rmine which use of the verb is c.ontcxtually appropriate. Elegant as it might seem, however, generalized recursive tyl)e resolution leads to conqmtational ineftlciency. Moreove.r, if wc ;Lssume that lexical entries are sort-resnlved during rule application, it is difficult, perhaps impossible, to avoid multiple solutions for an under.'q)ecifi('d lcxical item when its rule context Callnot lc~M to dete.rnlinistic disamhiguation. '\['his would be the case when parsing a verb such ~us brgtg with a noun phrase complement. As (:an he inferred with reference to the three |lS(:s of the w~'rb exemplilied in (2), three sohttinns are possible until either the subject or the next comphnnent is l)arsed:  (2) a. Mary brought l&amp;quot;ido h. Mary b~'(n/ght-171(7o to the party c. mary cookie  We trie!l t,o achieve ~t inore e(licient mid deterininis~ tic treatlnent by (h.weloping special-l)urpose facilities which make awfi\[able a guklcd approach the sort resohd,ion. The I)~si('. intuition underlying such an atteml)t is that for every class of lexical ambiguity there is a specitic word suhstructure whose instantiation is essential for disaml)iguation. For example, valency ambiguities lor verbs caa be generally resolved with reference to their complementation structure, ~s uoted above for the two uses of swim ill lid). Likewise, the ambiguity of nonfinals sueh as lamb which can be used as either simple nouns or m)un l)hr~scs in English (e.g. feed the htmb vs. cat lamb) can be contextually resolved with reference to dctc.rmiuer selection.</Paragraph>
    <Paragraph position="9"> We used i)rocedural attachments to rules to support contextmflly guided resolution of polymort)hic lexical type.s. The AI,E environment provides rather convenient facilitie:; to carry out this implementation iu the</Paragraph>
    <Paragraph position="11"> X is a typed feature structure (Carl)enter, 1992b:ch. 4).</Paragraph>
    <Paragraph position="12"> (3) a. list sub \[e_list, ne_list, comp_list .... \] .</Paragraph>
    <Paragraph position="13"> e_list sub \[\].</Paragraph>
    <Paragraph position="14"> ne_list sub \[ne_comp_list .... \] intro \[hd: bet, tl: list\] .</Paragraph>
    <Paragraph position="15"> b. member(X, hd:X) if true.</Paragraph>
    <Paragraph position="16"> member(X, tl:Xs) if member(X,Xs). Using the membership predicate above, we can define the ALE definite clause in (4) which would resolve polymorphic verb_synsem types by checking them against a list of unambiguous synsem types for consistency. null (4) solve_head_type(Lex_Type) if member (Lex Type, \[iv_undir_synsem, iv_obl_dir_synsem .... \] )  solve_head_type can be integrated with grammar rules as shown schematically in Fig 2 so that a verbal head exhibiting valency mnbiguity (e.g. iv_undlr_or_iv_obl_dlr~ynsem) with contextual instantiation of its list of complements - eAist or pp_comp-list, as defined in (5) -- would return a fully resolved FS (iv-synseln or iv_obl_synsem in Fig 1). This way of carrying out lexical type resolution has computationM overheads which tend to grow proportionally to the number of unambiguous lexical types. This is simply because lexical type resolution is done by unifying underspecified synsem FSs against a list of unambiguous lexical synsem FSs using the membership predicate: the longer the list, the heavier the computational overhead. With about thirty unambiguous verb types, we found that the disambignation of polymorphic lexical types using -~olve head_type with simple sentences was slower than enumeration of each distinct option through lexical disjunction--- although the difference in performance tended to couverge as we tried tinting longer and more complex sen.deg</Paragraph>
    <Paragraph position="18"> tl:e list\].</Paragraph>
    <Paragraph position="19"> Some improvements were obtained by eliminating the lnembership flmction and simply listing all possibilities as facts, e.g. solve~head_type(iv_undir_synsem) if true., solve~head_type(iv_obl_dir_synsem) if true. IIowever, we thought that better results yet could be achieved by exploiting conditions on constraint introduction rather than using unification with the list of unambiguous synsem FSs.</Paragraph>
    <Paragraph position="20"> Since in ALE path values can be introduced as constraints, an attribute and its value can be used to retrieve the type at which that value was introduced:</Paragraph>
    <Paragraph position="22"> Our basic idea was to define a rec.ursive definition of this facility and use it as a procedural attachment on rules to enhance lexical type resolution during language processing. For example, we could use the value for thc \[lead of tile compdist of a verb -- as provided  in tile course of rule application and the l)ath at which such value occurs to resolve the verb's lexieal type, e.g.</Paragraph>
    <Paragraph position="23"> \[ 7- rec restricts(iv_or iv_obl_eynsem, eyn : loc : comps : hd : pp_syns em, SubType).</Paragraph>
    <Paragraph position="24"> SubType = iv_obl eynsem This allowed us to carry out ambiguous lexieal type resolution without having to cheek type compatibility against a list of unambiguous lexicM types.</Paragraph>
    <Paragraph position="25"> We devised a version of rec_restricts which given an ambiguous lexical type and the resolving con|rain| returns the appropriate grounded type by I. retrieving all the minimal subtypes of the aml)iguous  type 2. collecting the constraints of eae.h subtyl)e into a list 3. returning the subtypes whose llst of constraints in- null clude tlm resolving constraint.</Paragraph>
    <Paragraph position="26"> The Prolog code lbr this Mgorithm is tus shown beh)w, where sub, intro and cone arc ALE predicates whicll encode subsumption, feature introduction and constraint declaration.</Paragraph>
    <Paragraph position="27">  arguments are: a (polyn,orphie) synsem type, and its resolving contraint a.s provided during the course of rule application, e.g.</Paragraph>
    <Paragraph position="28"> solve head_type( iv mldir_or iv obl dir_synsem, pp_synsem), In the compih*.d code for solve_head_type, the tmambiguous type given as output 1)y roe. restricts (e.g. iv_undir_synsem) is used to resolve, the input polymorl)hic type (iv_undir_or_iv_obl_dir_synsem) using unification of (atomic) synsc.m types rather than fldly tledged l!'Ss. This solution proved to t)e far more eltic|cut than the previous one. and never yielded worse results when compared to the enumeration of each distinct verl) valency option through lexical disjunctkm.</Paragraph>
  </Section>
  <Section position="5" start_page="698" end_page="699" type="metho">
    <SectionTitle>
3 Initial Results and Envisaged
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="698" end_page="699" type="sub_section">
      <SectionTitle>
Improvements
</SectionTitle>
      <Paragraph position="0"> Using the treatment outlined above, we have (leveloped a tyt)e lattice covering all ntajor comph!mentation patterns for English and (~erman (over 30 frames) with a variety of intermediate polymorphic types describing possible clusters of subcategorization ol)tions. At the same time, we have started to exploit the sltme technique R)r dealing with other cases of lexical alnbiguity, such tm the ability of noutiuals to functiolt as either nouns or noun l)hr;Lses, e.g. John drank beetle beer/beers/the beers, l'reliminary results are very en.</Paragraph>
      <Paragraph position="1"> c0uraging. For example a verb such ;m want whMt can be used as either a transitive (want a beer), subject equi (w.nt to .sleep) or object raising w'xb (want Mary to sleep) will only produce a single chart edge when followed by a VP complement, e.g.</Paragraph>
      <Paragraph position="2">  (6) 1% derivation(\[wmlt,to,sleep\]).</Paragraph>
      <Paragraph position="3"> 0 want 1 to 2 sleep 3 0 .........</Paragraph>
      <Paragraph position="4"> t .........</Paragraph>
      <Paragraph position="5"> &amp;quot;2. ....................</Paragraph>
      <Paragraph position="6"> 3 .........</Paragraph>
      <Paragraph position="7"> 4 ....................................</Paragraph>
      <Paragraph position="8">  With simple structures i~s the one in (6), the edwin|age ill using i)olynaorphic lexical types with sort resolution ~s comparexl to word sense euumeration by lexical disjnnction is minimal even though fewer chart edges are built. This is because there is a constant ow'Mlead when doing polymorphic type resolution through solve_head_type which in these c~mes is equivalent to building a l~w more lexieal edges. With more complex sentences, however, this overhead is soon offset, and the benefits of using lexical polymorphism t)eeome manifest, l,'or example, the analysis of a sentence like John likes that they want go come using polymort)hie verb types produce(l 23 edges and was about 15% faster than the analysis yMded using a lexicon with verb usage enumeration where 34 edges were built.</Paragraph>
      <Paragraph position="9"> We are also cent|dent that we can iml)rove the t)e&gt; formance of our el)preach in at, least two regards. First, we can reduce the co,nputational effort current;ly used in ensuring that the input lexieal tyl)e to solve_head_type has not been altered ~s a result of some previous rule el)plies|ion. Such a measure is needed, R)r example, when a w.'rb with l)olymorl)hie type undergoes morphological combination before the head-complement rule el)plies. In this clme, the semantics of the verb wouhl be altered with a consequent loss of the original (polymorphie) lexical type. This wouht make lexical type resohttion impossible. We must therefore avoid destructive ntodilications of the or|gluM lexical type while resohttion of such type is still possibh: by introducing in the sign a structure where the semantics of the bound morl)heme is stored until all verbal arguments are COltSllllled. 'l'he stored senl.atlties is then retrieved using procedural attach ments. This retrievM is eomputationally expensive im it is carried out by inealls of procedural attachme.nts, and we are now investigating the alternative of building the resulting semantics on line where it is currently stored.</Paragraph>
      <Paragraph position="10"> Second, we can make lexical type resolution by roe_restricts nlore deterministic in those cases where the solving constraint does not lead to a unique solution, as discussed earlier with reference to the verb bring. In the lexicon, briu9 is assigned the polymorl)hie. type tv_or_tv_obl_or..ditrans_syns(,.m which subsumes the tlu'ee uses of tim verb exemplified in (2): tv..synsem in (2a), tv._old__(litrans synsem in (2b), and ditrans syns(;m in lab). Because the three sub-types are consistent with a direct object subcategorizalion, z-ec_restr+-cts Callnot provide a unique solution when parsing bring with a nouu l)hrase complement.</Paragraph>
      <Paragraph position="11"> This is l)eeause, rec_restr+-cts carries out sort resolution of a I)olymorphic type by elmcking consistency of  the discriminating constraint against all minimal (most specific) subtypes of the polymorphie type. Consequently, tee_restricts would return three solutions for bring using the instantiation for the head of the compllst to np_synsem, as would the use of generalized recursive constraint resolution. In our approach, however, this inadequacy can be easily redressed by * changing rec_res~ricts so that sort resohltion is done by returning the maximal (least specific) sub-type of the input polymorphic type at which the discriminating constraint is introduced, and * modifying the grammar so as to support such a change. 1 As long as the same constraint is not introduced at several subtypes for each polymorphic type to be solved, these changes will ensure that sort resolution by tee_restricts is always deterministic.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>