<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-1067">
  <Title>An HPSG Parser Based on Description l,ogies*</Title>
  <Section position="3" start_page="0" end_page="412" type="metho">
    <SectionTitle>
2 ilasie Ideas
</SectionTitle>
    <Paragraph position="0"> An important distinction made in DI., but missing in traditionaI fealure formafisms, is the one between objects ;rod /)7)es. I)I. formulae eilher express that a type 1~ is more specific than (or subsumed by) a type t2 (tl :&lt; t2) or that an object o is an instance of a type or, using I)I, terminology, a conce U (o :: c).</Paragraph>
    <Paragraph position="1"> Applying Ibis schema to the task (5t&amp;quot; NI,P, we can say that the objects in this donaain are wnrds or phrases, and that the types are syntactic categories. I:urthermore, given a phrase el we have addiliomll relations between lhis l)hrase and its constitnents 02, 03, ..., nsually expressed as &amp;quot;02 is a daughter of ol&amp;quot;. In DL this ix modeled as '(or, 02):: dtrs', or equivalently as '01 :: dlrs:02'. 'dtrs' llltlS acts as a binary predicate or, using I)l. terminology, as a role. Nile thai roles can have more Ihan one value in contrast I(i fealures, which are funclional. We can thus write 'el :: dlrs:02 &amp; dtrs:03'.</Paragraph>
    <Paragraph position="2"> Note further Ih;lt the objects stand for occtJrrences of words or phrases, and thal different occtwrenees nf the slime word will be represenled by different objects. This is represented by writing 'o2 :: 15hen:or ', for example, It) express thai 02 is tin ocetlrrenee of Ihe form 'er'.</Paragraph>
    <Paragraph position="3"> This is all rather similar R) standard Ill ,.~(; nil;ilion, and the main difference is Ihat in addition to the feature structures used in IIIHG, we add ;ill additional level of objects, which we see as instances nf the feature strueha'es. Fea-Itlre structures Ihus correspond Io types or more precisely Io l)I. concepts. In a way, IJle objects ill 1)L are used to make the lIPS(\] feature slr(lchu'es\]~cr, v\[slcnt, i.e. 1(5 have pointers or names to refer to them.</Paragraph>
    <Paragraph position="4"> The additional level of objects allows a slraighlforward description of the parsing task. We start with a nttmber of objects, namely words, whose phonological value amd position ix known. We want to end up with a single object containing till these words as (not necessarily immediate) consthuents. Now the immediate dominance schemata in ill/ IIPSG loll us how 1(5 eonstHlct phrases from words or other phrases. Thus the main operation for building a phrase is to create a new object being an instance of an ID schema (note thilt it) schemata are feature slrtlctures and therefore concepts) and I(5 fill in the required daughters by us; W the objects available as building mate,'ial. This is achiewxt by choosing the 'functor'-daughter' and filling the required argumenls.</Paragraph>
    <Paragraph position="5">  Three p(lints are iinporlani in the following sot)ions:  1. Obviously, 0bjeels cannot be combined in a random way. In III'SG lhe ID schemata and Ihe lexieal onlrios eoniahl information concerning c()mbination wiih olher ptlrases. T will model lhis infornialion in DL tiiid rise standard I)1, itfferoneos lo cheek consisleilcy of eombinalions. Thus lho l)l, sysiom is used Io perform Ihe unifleaiion ilisk underlying Ill&gt;S(} and similar {Jnification Grainmars+ 2. All objecl can only be used )is building nlalorial \[+or a  phrase if il has ilol yel been used iis building maiorial for sortie olher phrase. I,'tlrlherinore, when looking, for dlitighlers o\[' li now l)hraso, we hay(:. Io fill lhoso daughiers for which a filler is required, but not yet specified, l will use tile cpi, vtemic o\]scr(t/or k Ill'()+ posed in IDonini el al. 921 Io formalize these inlu)lit)f is aud lhon use slandard Ill, rehieval for chocking these eonstrainls.</Paragraph>
    <Paragraph position="6"> 3+ lq)r synlaclically ambiguous expressions there is IIIoro ihan one possibility io coml/ino the words/phrases. Since the objects and especially ltiO rohitions between them fire viewed from differ eli( porspeclive,,i in lhe alloirlalivc' hlterprelali()ns, \vo lieed a nioehanisnt {ii Dl, to reproscnl lhose diffefonl views, i will i1sc xiluateU de.s'cri\]~tionx 'o :: c in s' in Iho followin,~ lo formalize lhis notion of differell( porsl)eClives+ Thert~ {s a l+O(ll~h corfespofldeiice boD.voon the silualiOllS iisod Io capture the specific in)eli)re(aliens and )he charls eroaled in charl pars.iBg. null</Paragraph>
  </Section>
  <Section position="4" start_page="412" end_page="412" type="metho">
    <SectionTitle>
3 The IhidcHylnl; Description i ,oEic
</SectionTitle>
    <Paragraph position="0"> \])eset'iptlon Logics Val y wit the lernl-buildhig operalor:l they contain+ In this set)ion 1 will pfesenl lhe ~yillax of file 1)1, which is u;ied hi ttio exainples given in Ihe next lwo seclions, l)ue to space lhnilali(ms I will li(tl specify lhe formal seinanlics for this 1)1. (see, for exalilo ph', \[I\[oppe et al. 93, Quanlz, Schlnilz 9&lt;!1 for a model-</Paragraph>
    <Paragraph position="2"> When specifying file fralgmerlt and the pars;el + in the noxl ,';(+clion.+ I will use a ilolallllil based Oil Ilie PROI.()G inlerface pl.,')vide,:t lly the BACK sysIetn lll.l)pe et al. 931. hi BACK a dislinclion is lllade belween lerm il~ti'oduclion,'-; or dofinilions, and conslritlnl,.liko rlllCS. A lernl ilaino can Do inlroduccd cii}ler ilS/~/'/,'lli{il,t&amp;quot; (1, :&lt; l), i,e. ()lily lleCeSSilfy (:ondilions are t~,ivon, or as (Iothlod (l, :: : t), i+e. necessary arid sufficient condil{oils fire (,iron. A i'ulc Cl &gt; (?2 iliCatlS lhal each objocl being at) instal)co of el {s a\],~;o {ill insl;ili('e of (;2.</Paragraph>
    <Paragraph position="3"> &amp;quot;lhe foraulhl 'extend sil(sl,s;0' exprc.sses lhe (act thai siltlali(m se is ~lri extensicm of siltiali(in sl. This mean:-; Iha\[ 'o :: C in sl' implies 'o :: c in ~;7.' for all objeels tl and colicopls e.</Paragraph>
    <Paragraph position="4"> It(1 order to dislitiguish belwcen tcllilly, and quelying itfformaiioli I will llSO 'o ~: (; ill s' l:or tolls and 'o '?: ill s' for qilOi+\[e~. I ftirlhoHnore aSstlnie lhlit ii toll only suceeds if it is consistent with lhe previously enlered informal)on; otherwise it ill)Is. When the object used in a query is a variable, the syslem will relrieve all known instances of a concel)l, i.e. 'Object +?: in s' will reltirrl the objects known lo be inslances of 'e' in 's' by backtracking.</Paragraph>
    <Paragraph position="5"> Note |ha |the epislemie operator k will ()lily be used in queries. It can thereR)re be straightforwardly integrated into exisling I)1. syslems. Since this is also true for situated descripliorls, lhe parse,&amp;quot; presented in Seelion 5 is largely b~lsed Oil stil\[ldilrd inference capabilities of DI, systems.</Paragraph>
  </Section>
  <Section position="5" start_page="412" end_page="413" type="metho">
    <SectionTitle>
4 A Snlall l:ragment
</SectionTitle>
    <Paragraph position="0"> ll'~ lhis section }\[ will present examples from an In .'S(;-slyle fragment for German modeled in DL. Due to space limilalions I will nol specify all the information contained ill Ihis modeling but only lhe one needed 1o illustrale the main characterislies of file formalization and lhe example sentence 'Die sch6ne l:rau sieht sic' discussed in Ihe nexl seclion.</Paragraph>
    <Paragraph position="1"> The fragment is based on Ihe presenlalion in \[PcJlhtrd, Sag871 and ils applicalion to German in \[llill191L A main difference between lny I)I. modelillg and slandard Ill&gt;S(} modeling is lh;l |\]\[ ;IVOi(l fealure imthes which would inlroduce st;i)erfluous ITll+ objects. There is thus tit) feature 'head' in my modeling since it would yield the im,+oduelion of head objects whose ontological status seems controversial. Consequently, my IIead Fealure Principle specifies cqtdwdence not for a shigle fealure 'head', but rather for each head feature separately.</Paragraph>
    <Paragraph position="2"> The fragmenl eotlIilii/s five main categories, llamely tie(m, ,'lfl, i,e,'D, de/, and tu//. For ilhlslrallon, the definili(ms  ()i II()llll and ,qt'J afe j;iven l)elow: II(l{lll : -- lll~lj:ll &amp; lex: I&amp;quot; &amp;quot;p : : maj:u &amp; lex: Phrase slrtlcltlre is represented by roles as Ihe I'ollowiIig: till.'-: :-&lt; dom{lln(si/~il) t~ rallll~o(sl~l,tl) COOli)_ dlts :&lt;5 dll.'; COlllp dh'\[ :&lt;5 coinp dim &amp; foal )lead dlr :&lt; dh~ &amp; teal funcior_dir :&lt; dim &amp; feat  The fealut'e ' funcior, dl r' will be used by lhe parser it) st)coify Iho sib, n aclin{ ~, as funelor of a now phrase, lls wihle will be idenlical to Iho vahio of 'head. dlr', 'adj.dh&amp;quot;, ()t 'filler dh&amp;quot;, dependilit; I)ll Ihe p',lrlietllar/##tnlc'(l/{lle Z,)on//)la,'sc(: (It)) sehemli ilSOd. Nolo lhtil Iho daughlors which ;ire modehxl as feilltlres ltzt2 funcliomil, i.e. no phrase can have lwo fillcr:~ for 'heild dh&amp;quot;.</Paragraph>
    <Paragraph position="3"> (;orfest)ondinl_; {o these dau{,Jiler roles and foal(ires we have art, umonl roles :lnd fealuros a.&lt;; 'comp_argl' elc, l</Paragraph>
    <Paragraph position="5"> Thus hi a 'colnp_sll'tlCItlrt;' Ihe 'head dlr' ,'lOiS as a fililClOr.</Paragraph>
    <Paragraph position="6"> Nolo lhal il has It) be cxplieilly slalod wholher a eerlain \['eahlre is enlply, e.g. 'lie(adj..all()' for 'comp_slrueliiro'.</Paragraph>
    <Paragraph position="7"> Ill, syslcms ilSS(llile all ()pell world aild \[like all descrit)iions  as being partial, i.e. the fact Ihat there is cutxently no known filler for a role at an object does not imply thai there will never be one.</Paragraph>
    <Paragraph position="8"> &amp;quot;Ilm fragment contains six ID schemata, namely three for noun phrases, one for verb phrases, one for adjtmets, and one for topicalization.</Paragraph>
    <Paragraph position="10"> For tile lexieal entries I will use three moq)ho-symtactie features (n form, case, gen) to illustrate agreement between nouns, adjectives, and determiners. Agreement concernins ease and gender between nouns and detenniners is modeled by specifying that tbe value of tile feature 'case' at a common noun is Iho same as the vahle of Iho featnre 'ease' at the object filing the feature 'eomp_argl' (which is the determiner).</Paragraph>
    <Paragraph position="11"> Below are lexical entries for 'frau' and 'sic':</Paragraph>
    <Paragraph position="13"> Note the bierardfieal nature of the modeling--the sub-categorization information is specified for common nouns and pronouns in general, and is then inherited by each specific common noun and pror~oun. Information shared by all forms of a lexeme is specified as a property of Ihe lexeme, whereas information specilic to a parlicuhw form is specified for Ibis form only.</Paragraph>
    <Paragraph position="14"> Adjectives require non-saturated noun phrases as arguments and agree with them wrt ease .'rod gender:</Paragraph>
    <Paragraph position="16"> Finally, the lexieal enlry for 'sieht':</Paragraph>
    <Paragraph position="18"> Note that for verbs taking more than two arguments we need addilional features 'comp_arg3' and 'comp_arg4'.</Paragraph>
    <Paragraph position="19"> In addition to tile information modeled so far we need a formalization of the principles underlying the combination of signs in flING. Some of these principles hold only for ph,'ases and not for signs in general. A ph,'ase is defined as follows:  &amp;quot;l'\]m parsing process presented in the next section is essentially triggered by signs which can act as functors, namely signs wilh unsaturated subeat \[isis, signs wilh slashes, and pronouns:</Paragraph>
    <Paragraph position="21"/>
  </Section>
  <Section position="6" start_page="413" end_page="415" type="metho">
    <SectionTitle>
5 I)L-Based Parsing
</SectionTitle>
    <Paragraph position="0"> In Ibis section 1 will present the basic structure of a l)L-based parser for tile above fragment. The parser is realized by five main predicates. I assume that tile initial informalion given In the parser consisls of descriptions of the words occurring in tile expression to be parsed. Consider the ambiguous sentence  (1) Die sch/3ne Frau sieht sie.</Paragraph>
    <Paragraph position="1"> (2) &amp;quot;llle pretty woman sees her.</Paragraph>
    <Paragraph position="2"> (3) Tlm pretty woman she sees.</Paragraph>
    <Paragraph position="3"> &amp;quot;llm initial DI. representation of this sentence is:</Paragraph>
    <Paragraph position="5"> Given Ibis information the parser builds phrases from tile live words. This is done by creating new phntses until no more combim~tions of signs are possible. &amp;quot;l\]m parsing succeeds if the words have been all used up and a single phrase results:</Paragraph>
    <Paragraph position="7"> pai~e_sign (NcwSit,FinSil).</Paragraph>
    <Paragraph position="8"> Note that tile epistemic. C/'oncepl 'no(k(inv(dtrs)))' is used to determine whether a sign is still available for phrase buiMing. An object is an instance of this concept if it is nol a liller of some 'dtrs' role at any other object.</Paragraph>
    <Paragraph position="9"> Tim basic idea of building a new phrase is to look for a sign which can act as a funelor, to choose an ID schema in which lifts sign is a functor, and to find the required ;trguments of tile functor. Finally, the linear precedence rules are checked and, if necessary, traces are introduced. 1</Paragraph>
    <Paragraph position="11"> Sign ?: funelor &amp; no(k(inv(dl~,'s))) in Sit, selecLid_schem a (S ign,S it,Piu-ase,NewS it), eomplete~lrguments (S i gn,NewS it,Nextgit), check J ps_and_continuity (Phrase,Sit,NextSit,FinSit).  Seleetkm of an II) schema is realized in a rather naive and simple way---we just take an ID schema and try to create a new phrase its an instance of lit is schema, where the feature 'ft,nctorxllr' is filled by tile funclor.</Paragraph>
    <Paragraph position="12"> IDue to space limitations I do not specify Ihe predieale 'checkdps_and_continuily' in this paper</Paragraph>
    <Paragraph position="14"/>
    <Paragraph position="16"> Information about existing II) schem;fla thus has to be encoded as facts of the form 'kt_sclmma(idl)', elc. Tile predicate 'extend_sit(Sit,NewS\]t)' ix used to tell tile DI+ system to create a new situation which is an extension of tile current situation.</Paragraph>
    <Paragraph position="17"> Note that no further knowledge al)out tile ,'tctual roodcling of It) schemata is uscxl in tile parser except for tile fact that each ID schema has a 'funclor_dtr'. Note furlher that the DL tell will fail if tile information known about the ftmetor cannot be unilied wilh the information required by tile ID schema for the filler of 'funetc, r_dtr'.</Paragraph>
    <Paragraph position="18"> In order to complete tile arguments of the functor, tile parser cheeks for each argument feature ArgFeat whclhcr an argument is required (somc(ArgFeat)) but not yet specified (no(k(ArgFeat))). If so, 'lind_arg' looks for such an argument and enters it as filler for ArgFeal. Then tile remaining arguments are completed.</Paragraph>
    <Paragraph position="19"> complete.argumenls (Functc, r,S il,FinS i1) :arg_feature (AtgFeat), Functor ?: some(ArgFeat) &amp; no(k(ArgFeal)) in Sil, !,tind_mg(Functoe; Sit,ArgFeal,NewSil), complete_arguments(l:unctonNewSiI,HnSil).</Paragraph>
    <Paragraph position="20"> complete_argumenls (_,S it,Sit).</Paragraph>
    <Paragraph position="21"> Again we need to introduce facts sl)ecifying tile arglnnents used in tile fragment, e.g. 'arg_fealu re(comp_arg I ) '. If an argument is rc*qt, ired it has lo be filled, therefore tile Cut, Thus tile recurs\]on lean\]nares successfully only when all required arguments are actually tilled. Nol0 thai the only information about argument structure needed t)y the parser are facts of tim form 'arg_feature(comp_argl)' for all argument features.</Paragraph>
    <Paragraph position="22"> To find an argument the parse,&amp;quot; looks for a sign which has not yet been used for phrase building and ,asserts it as filler for Ihe argument feature. Again, if unification is not possible due to conflicting constraints (e.g. agreemenl), the DL tell will fail.</Paragraph>
    <Paragraph position="23">  tind_arg(Funclot;Sil,Argl:eat,FinSit) :..</Paragraph>
    <Paragraph position="24"> new_ ph rase(Sit.NewSi,), find_a~g(l&amp;quot;unctcw,Sil.ArI,Feat,lqnSil).</Paragraph>
    <Paragraph position="25"> The second clause is needed Io create a required ,'lrgument which has not yet been build Ul). Ill this case 'new. phrase' is called It) Creale a new potential ;,rgumellt.</Paragraph>
    <Paragraph position="26"> For the sentence 'Die sch6nc Frau sieht sic' we obtain two different parses, since bolh 'die schoene frau' and 'sic' are ambiguotls between nolni/lalive ;ilia accusative ease. &amp;quot;llle reading according t() which 'die schoene frm,' is subject is shown in Figure I as a t)hrase slructure tree. Some of the eorresporvJing infomullion conlained in Ihe I)I. situation represenling this reading is given below:</Paragraph>
    <Paragraph position="28"> In the second purse tit and P9 swap places, i.e. l)9 is the 'comp.dtrl' of \])12 and Ill is the 'comp.Atr2'.</Paragraph>
    <Paragraph position="29"> The rest,It of the parsing process illustrates tile objectcenteredness of D\]. representations. The constituents of tile ulterance are explicitly modeled and can be used for extracting or specifying further information. Thus we can choose to introduce at feature 'subject' and add tile fact '1)1.1 :: subjecl:ps', or we can retrieve all tile salurated noun l)hrases (Phrase 2: n t&gt; &amp; no(args)).</Paragraph>
    <Paragraph position="30"> &amp;quot;It,is object-cenleredness is useful fo,&amp;quot; disambigualion, for example for a,laphor\[i resolution, as ilhtshated in \[Quantz, Schmitz 941.</Paragraph>
  </Section>
  <Section position="7" start_page="415" end_page="415" type="metho">
    <SectionTitle>
6 Interpretation as Exception Minimization
, L *
</SectionTitle>
    <Paragraph position="0"> I will now briefly sketch how ihe parser presented in the previous section can be extended to perform disambiguation by exception minimization as proposed in \[Quantz 93\]. In case of ambiguous expressions tile parser will return more than one situation. Tile basic idea of interpretation as exception minimization is to model additional preference rules nee, ded for disambignalion as DL defaults, and to choose the inteqlretation violating a qualilatively minimal set of defaults.</Paragraph>
    <Paragraph position="1"> A Preferential Default Description Logic (PDDL) based on weigthed defaults has been developed in \[Quantz, Ryan 93\]. A weigthed default /5 has the form cl &amp;quot;-*,~ ca, where cl is called the premise of/5 (/51,), c2 the conclusions of 8 (~5~) and n the weight of/5 (w(8))--the higher the weight, the more relevant tire default. For lhis PDI)L a formally well-behaved preferential entailment relation o ~&gt;2 is presented, which is based on an ordering on DL models deg\[5~:. The basic idea of this preferential semantics is to compute a score far each model by summing up the exceptions to the defaults. Models with lower score are then preferred because they qualilatively minimize the exceptions. It is straighlforward to carry the idea of scoring and ordering over from models to situation. To do so, we compute for each situation s and each default 6 tile exceptions--those objects for which 'Object ?: 6~, in s' sncceeds and 'Object ?: /5~ in s' fails.</Paragraph>
    <Paragraph position="2"> If there are several possible interprelations for an expression we choose the interpretation given by lhe situation with the lowest score. (Note that there may be Iruely ambiguous expressions which yield situations with identical scores.) Thus taking the above example, we might use a preference for topical ization of subjects to prefer Ihe parse shown in Figure 1. This can be achieved by simply introducing a defat, lt np&amp; top:+ ~5 case:nora Obviously, this default is a rather weak one and can be overwritten by information stemming from selectional restrictions \[Schmitz, Quantz 93\].</Paragraph>
    <Paragraph position="3"> In principle, it is possible to use preferences stemming from weighted defaults already in tile parsing process--situations whose score is higher than a specilied threshold are not processed any further. Thus instead of producing all parses in tile first step and ordering them in a second step, the parser would Ihen only produce tire preferred reading.</Paragraph>
  </Section>
class="xml-element"></Paper>