File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/86/c86-1005_metho.xml
Size: 24,546 bytes
Last Modified: 2025-10-06 14:11:48
<?xml version="1.0" standalone="yes"?> <Paper uid="C86-1005"> <Title>EX'I~NDING Tfiq~ PlKPRESSIVE CAPACITY OF Tf~E SFIMANTIC (DMPONENT ~F ~\[~, OPFJ~A SYST~</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> EX'I~NDING Tfiq~ PlKPRESSIVE CAPACITY OF Tf~E SFIMANTIC (DMPONENT ~F ~\[~, OPFJ~A SYST~ CELESTIN SEDOGBO CENTRE DE I~ECH~RCHE BULL </SectionTitle> <Paragraph position="0"> OPERA is a natura\], language question answering syst~n allowing the interrogation of a (hta base consisting of an extensive listing of operas. The linguistic front-~d of OPERA is a c~nprehensive grm~r of Vrench, and \]1:s semantic c~nponent translates the syn* tactic analysis into logical formu\]as (first order \]o-. gic forrnulas) ..</Paragraph> <Paragraph position="1"> However there are quite a flew constructions which can h~ analysed syntactically in the grmn\[mr }ant for which we are l\[~able to specify trans\]ations, t'~orcmlost muong these are m)aphoric and elliptic contructions.</Paragraph> <Paragraph position="2"> Thus this \[kqper descri\[~s the extension of! OPERA to anaphoric and elliptic c~mstructions on the basis of the Discourse Representation Th\[~)ry (DRT).</Paragraph> <Paragraph position="3"> l- \]INTRODUCTION OPEPJ~ is a natural language question answering system al\].owing the interrogation of a data \[mse consisting of an extensive listing of operas and their ass(x~\]ated caracteristics (c(~ll\[~sets~ time of ccm\]pos/.tJon v cecor-. dings, etc .... ). The \].inguishic front.-.~md o\[ OPF, I~A i.s a c<~nprehensJve grat~nar of Fwench based essentially on the string gra~ar for~m\].ism (cf for exc~n\[)le \[SAGE\[{~S1\] o~: \[SALKOFF~ 73\]); the details of the syn-~ tax used in OPE|{A are descri\[x:d in \[SEIIYOGBO, 85 \]. \[in \['his paper we shall descri\[x~&quot; Lhe scmmrrh:i.c c~u\[mnent ol OPERA as it stands now. In addition we shall d:\[scuss a nun~x~,r <)if prob\].c:~natic constructions which are not handled at present in the systc{~ and we outline a 17os-.</Paragraph> <Paragraph position="4"> sible solution for thon in the setting of this systemdeg It has bc~coale more ~/ld nDre C\]l ~k~\] \[ that the n~ost natural and transparent fon~ulation of the s~imntics of natural languages can be constructed in terms of a lo.gical so~Itics i.e a sel~%ntics b~sed on the usual mode\] theoretic interpretation of predicate logic.</Paragraph> <Paragraph position="5"> This was in fact already realized as early as 1973 by A. COI~4~:RAUER and his group ( CF \[COLMF/~AUER & ai,73\], \[PASERO, 73\], \[COI~MERAUER, 79\]). In these papers the sexnantics for natural languages is specified in terms of the truth conditions of predicate logic formulas into which these utterances are translated in a syst~natic way. ( It should be n~ntioned that in many respects the way these translations are obtai ned as well as the translations the~mselves resemble the even earlier proposals of MONTAGUE and his students (cf \[MONTAGUE, 74\]). The work done in COLMERAUER's group has however the additional merit that these proposals are presented in operative syst~ns for nmn-machino dialogues. Indeed in these early papers we find sketched two alternatives for the systs~mtic manipulation of the translations provided for natural language discourse: This research was supported in part by ESPRIT, in the ACORD project(P393).</Paragraph> <Paragraph position="6"> a) given a data base in relational (e.g Prolog like ) form the answers to queries (which are themselves translated into logical forms) ~n he given in terms of the satisfaction conditions of these formulas. I.e a query like &quot;does every student own a hook? &quot; will receive the answer &quot;yes&quot; if its translation &quot; for a\].l x (student(x)-> there exists y ( book(y)& own(x,y))) &quot; is true in the data ~ise. Here the answers to queries are obtained by simply evaluating the translation in the data base (or mode\].)deg b) another approach however consists in regarding the data base as a set of fomaulas and the process of answering a query as a deductiondeg A query wil\] be said to be true with respect to the ~ita \]mse if and only if it can I~ <\]educed frc~n \]to (Naturally as in the case of the first approach, this meth(x\] app\].ies (~qual\]y well to closed sentences ( idege yes-no questions) as well as to open sentences ( i.e W\[I-questions)deg \],'or an application of this im~thod as wo\].l as for disc:ussion of its advantages (cf \[PASERO~ 73\])deg \].'o~. reasons which w~ sha\]l not spell out here we have chosen the ~ode\] the(met\]ca\] approach (\].e the first approach ) in OPERAdeg Needless to say it: is not Loo difficult and in fact, sclnetimes necessary~ as we shall see below, to c<mlpl6\]nent the pure se~nantic evaluation with deductivo capacities..</Paragraph> </Section> <Section position="2" start_page="0" end_page="23" type="metho"> <SectionTitle> 2- Sh:MANTICS VIA TRANS~TION IN'IY) LOGIC </SectionTitle> <Paragraph position="0"> Even if it is clear in nmst respects how to obtain predicate logic: translaticms for a large w~riety of natura\] language sentences, cf \[WAR\[~FM & PEREIRA, 82\] we repeat here :\[or the sake of clarity the essentials of the translation process in OPERA.</Paragraph> <Paragraph position="1"> Consider for c~ample the following sentences : (a) Berg a c<~npos6 un c~era en trois acres (Being composed an opera in three acts) (b) aucun c~upositeur gl.l~mnd n'a compos4 un opera en 1912 (no german cc~poser composed an opera in 1912) (c) Chaque opera de Berg a dt@ ~\]registrd par Karajan every opera by Berg was recordcw\] bY I~rajan. The syntax of ~?ERA (cf \[SEDOGBO, 85\]) yields syntactic analysis as in (d).</Paragraph> <Paragraph position="2"> These trees are then translated into the following predicate logic formulas :</Paragraph> <Paragraph position="4"> The evaluation component of OPERA then evaluates these formulas in the data base. Quite a few syntactically complex queries can thus be formulated in this system.</Paragraph> </Section> <Section position="3" start_page="23" end_page="25" type="metho"> <SectionTitle> 3- FROM TRANSLATIONS TO DRSs </SectionTitle> <Paragraph position="0"> OPERA was conceived not so much as an effort towards an implementation of viable integrated natural language question answering system, as a testing ground for a rather extensive fragment of French. In fact many constructions ( in particular cc~plement constructions ) which are accemodated in the syntax fragn~nt of OPERA are such that they cannot effectively be applied in the setting of OPERA's data base. H(~ever there are quite a few constructions which can be analysed syntactically in the gra~ir but for which we are unable to specify translations, even when their intended translations are essentially first-order logic formulas. Foremost among these are anaphoric and elliptical contructions. It is clear' that any c~nprehensive treatment of such phenomena cannot restrict itself to the analysis of isolated sentences i.e to isolated queries. For what could a sentence like: when did he compose it ? mean ? or what query could be expressed by the phrase &quot;and Britten?&quot;? There is not much that these sentences could convey in isolation. But consider the following dialogue : Question: did Beethoven cazpose Fidelio ? Answer : yes.</Paragraph> <Paragraph position="1"> Question: when did he ccnkoose it? A/iswer: in 1812 Or consider another very natural interaction : Question: how many operas did Beethoven compose? answer: 1 Question: and Mozart? Answer : 15 It is clear that an adequate account of anaphoric and elliptic constructions must at least take into account the current situation of the dialogue. How to define the notion &quot; current situation&quot; is by no means a trivial task. For one thing we cannot simply asssume that the current situation is identical to the sequence of the question-answer pairs that make up the dialogue. Much more is involved.</Paragraph> <Section position="1" start_page="23" end_page="25" type="sub_section"> <SectionTitle> 3.1 TOWARDS A TREATMENT OF ANAPHORA </SectionTitle> <Paragraph position="0"> Even though pronouns raise no probl~u at all as for ~ as far as t~ syntax is concerned, no one has been able to provide a unified and systematic account of anaphoric linking. As recent work has shown one must distinguish on the one hand various types of pronc~ihal reference and on the other hand show how the resolution of anaphora must appeal to common Lmderlying mechanisms.</Paragraph> <Paragraph position="1"> It is one of the achivements in \[KAMP, 81\] to show the way towards that unified treatment .</Paragraph> <Paragraph position="2"> In fact we shall base our proposals for beth the treatment of m\]aphora and of ellipsis on the concepts and technics introduced by KAMP in his theory of discourse representation.</Paragraph> <Paragraph position="3"> Since the current translation cx~mpenent of OPERA does not take into account pronouns at all we must first provide an extension which is also able to deal with pronominal referemce. Syntactically pronouns can occur in all NP positions characterized bY the grammardeg We shall 1~mrk each occ\[mence of a pronoun by indicating its gender/number features; As a first step in the translation procedure we shall therefore simply take these translations to he identical with themselves. Thus the translation of sentences like : Berg l'a composd (Berg c~mpesed it) will be : co, loose(Berg, \[it,null,sing\]) similarly a sentence like : chaque cc~npositeur qui a compos4 un opera l'a enregistr4 null (every ccTmposer who wrote an opera recorded it). will be translated into : chaque(x,exist(y,opera(y)&composer(x)&c~npose(x,y)), record(x,\[it,null,sing\]))) We shall call translations containing occurrences of pronouns &quot;unresolved translations&quot;. Notice that contrary to the practice in \[MONTAGUE, 74\] we do not index pronouns either in the syntax or the initial tranlation phase. The resolution of anaphora will take as input &quot;unresolve<\] translations&quot; and yield as output a rather different type of s~nantic representation namely a discourse representation structure (DRS).</Paragraph> <Paragraph position="4"> What is DRS? In general we shall say that a DRS is pair consisting of a (possibly empty) domain of discourse referents U and a set of conditions CON. We shall for the present discussion take into consideration only three types of conditions : a) atc~nic conditions, which consists of n-ary predicate P and terms.</Paragraph> <Paragraph position="5"> a term is either a discourse referent or a proper name; among the predicates we single out the eqdality predicate &quot;=&quot;. b) conditional conditions which have the form =>(kl,k2) where kl and K2 are again DRSs.</Paragraph> <Paragraph position="6"> c) negative conditinal conditions which has the form #>(Kl,k2).</Paragraph> <Paragraph position="7"> In a more comprehensive treahnent Jt is clear that we shall need further types of conditions.</Paragraph> <Paragraph position="8"> Thus for the final translations of unresolved translation we want to arrive at DRSs like the following :</Paragraph> <Paragraph position="10"> There are precise truth definitions for DRSs which we shall not s\[~ll out here however (cf KAMP for details). In any case the first DRS is logically equivalent to the formula : compose(berg, loulou) and the second DRS is logically equivalent to the predicat lo-</Paragraph> <Paragraph position="12"> What is interesting in the second example is of course a fact that the pronoun &quot;it&quot; has as its syntactic antecedent the noun phrase &quot;an opera &quot;. The se~qntic force of this noun phrase cannot however be rendered in terms of an existential quantifier at \].east not if we want to establish an ~maphoric \]ink between the existential quantifier and the variable representing the pronoun in the consequent of the conditional.</Paragraph> <Paragraph position="13"> MO~I~AGUE for instance runs into this proble{n in Vi~ where the pronoun is either left unbound or either is bound by the existential quantifier when the latter occur outsi~ the conditional all together. This wide scope reading of the indefinite NP obviously gives them the wr(mg interpretation. As does of course the translation leaving the pronoun unbound. In the DRS on the other hand we get a universal reading for an opera as the result of the interpretation of the conditional. null Even though inside the antecedent of the conditional we treat the occurrence of the indefinite noun phrase &quot;an opera &quot; as we would in an ordinary indicative sentence, where of course ~le interpretation of the indefinite article corresponds more naturally to an existential quantifier. This is one of the features of IQNMP theory that we shall tmke advantages of in the present proposal.</Paragraph> <Paragraph position="14"> As we said above we shall transform unresolved translations stepwise into DRSs. Needless to say we could of coucse set up the translation procedure in such a way that we obtain DRSs directly from the output of the syntactic component. But this would entail major revision of the entire translation algorithn in any event the way we propose to derive DRSs can he regarded as DRS construction algorithm in its own right. On this view &quot;our unre~uced translations&quot; play the role of an intermediate structure between syntax and DRSso Assume that t is an unresolved ~:anslation of a sentence . We first generate K(t) the discourse representation structure corres~:mding to t, in the following way : if t is a univemsal tree i.e a formula of the fon~ Let t be a tree with the determiner &quot;aucun&quot; as its top node, i.e a formula of the form : aucun---x ---fl .... f2 we procede as in the case above except that the condition we emter into K(t) is now a negative universal condition, i.e a condition of the form =>(kl,K2).</Paragraph> <Paragraph position="15"> Suppose t is dallinated by exist i.e t has the form : exist .... x .... fl we create a DRS K(t) whose domain contains x and we add the result of traulsforming fl as the conditions to K(t).</Paragraph> <Paragraph position="16"> Suppose t is a tree dc~inated by &quot;et&quot; of the form : et---fl ---f2 i.e a formula we create a DRS K(t) with an empty domain whose conditions are the result of transforming fl and f2 with respect to K(t).</Paragraph> <Paragraph position="17"> Finaly suppose the unresolved translation is non quantifie~\] then we enter it as is into K and we add all occurrences of proper names into the domain of K. This provides the induction basis for the tranformation. Let K' k~ a DRS with dc~min U' and conditions CON' and let t he a tree i.e a formula. The result of transforming t in K' is the application of the above three rules to t. When we no longer have any tree to process all conditions in the principle K i.e the DRS representing the sentence to be transformed, as well as all conditions occurring in the sub-DRSs of K will now be conditions in the language of DRSs or atomic conditions containing occurrences of pronouns.</Paragraph> <Paragraph position="18"> How are these be eliminated ? Let us first consider an example the sentence &quot;every composer dedicated an opera to a conductor that he has admired&quot; has as its unresolved translation:</Paragraph> <Paragraph position="20"> We indicate the contruction of the DRS stepwise</Paragraph> <Paragraph position="22"> We will give below the exact method for translating formulas of OPERA into DRSs.</Paragraph> <Paragraph position="23"> One of the most important features of DR-theory is the precise constraints on the antecedents of pronouns.</Paragraph> <Paragraph position="24"> Let K' be a DRS embedded in K (which is a DRS too); the antecedent of a pronoun occurring in K' is the list consisting of the union of U(K) and U(K') if and only if K is accessible to K'.</Paragraph> <Paragraph position="25"> For a precise definition of the notion of accessibility cf \[GUENTHNER & I~HMAN,85\].</Paragraph> <Paragraph position="26"> Let us now illustrate the notion of accessibility by giving the table of accessibility of the DRS above: ~</Paragraph> <Paragraph position="28"> By the transitive closhre of the accessibility relation we obtain for example all the possible antecedents of a pronoun occurring in K2 (e.g a pronoun occurring in K1 cannot have as antecedent a referent of K2).</Paragraph> <Paragraph position="29"> For the clarity of what will follow, we will call &quot;unresolved predicate&quot; (abreviated UP) a predicate whose arguments included at least one pronoun. Then to resolve an UP, one must replace the pronoun arguments with appropriate referents accessible.</Paragraph> <Paragraph position="30"> The idea is to transport during the translation of formulas, a list of antecedents accessible according to DRSs constraintsdeg How is this list to be constructed? As shown in (3.1) a universal tree is translated into a DRS K= =>(KI,K2); the antecedent list of K which we note L(K) is 6{npty.</Paragraph> <Paragraph position="31"> The antecedent list of K\] is L(K1) and L(K2) = U2 + L(K1) (we denote the union of sets by the symbol +). The existential tree is translated into K = \[U,CON\] and the antecedent list L(K)=U.</Paragraph> <Paragraph position="32"> Let f be a formula with &quot;aucun&quot; at its top node; f is translated into the DRS K ~>(KI,K2). The list of antecedents L(K) of K is empty and L(KI) is U1 and L(K2) =U2 + L(KI).</Paragraph> <Paragraph position="33"> Let us call a DRS containing &quot;unresolved predicates&quot; an &quot;unresolved DRS&quot;. Thus each unresloved DRS is a pair of the form K = \[K,L(K)\] where L(K) is the antecedent list of the DRS K.</Paragraph> <Paragraph position="34"> To resolve an unresolved DRS, each unresolved predicate (UP) must be resolved according to the following mule : for an unresolved DRS K = \[K,L(K)\] with K=\[U,CON\] a UP P of CON is transformed into the logical predicate P&quot;which is obtained by unifying pronoun arguments of P ih L(K).</Paragraph> <Paragraph position="35"> * The example &quot;every composer dedicated an opera to a conducter that he has admired&quot; treated in session 3.1 will illustrate how unresolved DRSs are resolved.</Paragraph> <Paragraph position="36"> Given the unresolved DRS K = \[K,nil\]</Paragraph> <Paragraph position="38"> K will he resolved by application of the rule described above, i.e: The pronoun he is \[mifiable in the list (x,y) of antecedents, to z. Therefore the unresolved predicate will be translated into a~idre(x,z).</Paragraph> </Section> </Section> <Section position="4" start_page="25" end_page="25" type="metho"> <SectionTitle> 4- EX'I~NDING THE DIALOGUE CAPACITY OF OPERA </SectionTitle> <Paragraph position="0"> A proble~ arises in interrogating a data base in natura\]. language which is the problem of dialogue situation. For example : (a) qui a c~t~posd Loulou (who composed \[mulou ?) (b) oh est-il n4? (where is he born?) This dialogue can he translated into a DRS containing the semantic representation of the two sentences. But if there is no interaction with the data base, the anaphoric link of the pronoun in (b) will be an unbound variable and not for example the individual &quot;Berg&quot; (as Jt is Berg who c~nposedLoulou). What we prefer for a question-answering system is to take into account the situation of a question and its answer.</Paragraph> <Paragraph position="1"> A possible solution could be. to consider that after the eval\[mtion of a query, its c~rresponding DRS is therefore instantiated (i.e. its unbound variables are now bound); in such a situation we loose the reading of the DRS as a logical formula and moreover we cannot represent the instantiation of a splitted DRS (e.g. a universal formula).</Paragraph> <Paragraph position="2"> 4deg1 q~E NCZI'ION OF A &quot;CURRENT SITUATION DRS&quot; The solution we proposed is to separate the DRS of a formula from the DRS for the situation.</Paragraph> <Paragraph position="3"> Let t be a formula and K(t) :\[U,CON\] its translation into a DRS.</Paragraph> <Paragraph position="4"> The DRS of situation is the ORS KSS(n) = \[USS,CONSS\] with USS containing the instantiation of referents of U, CONSS (~pty, and n denoting the current state of the dialogue (i.e. the occurrence of the question during the dialogue).</Paragraph> <Paragraph position="5"> The rule for UP resolution presented in the last session, must now be modified in the following way: let n be the current state of the dialogue and K be an unresol~<\] DRS; the antecedent of a pronoun occurring in K is containel in the list consisting of the concatenati(xl of L(K) and L(KSS(n-I)).</Paragraph> <Paragraph position="6"> The antecedent list of KSS(i) is defined in the same way as that a noraml O~5, i.e. L(KSS(i))=USS(i). To illustrate what is abever we will treat the dialogue above:</Paragraph> <Paragraph position="8"> To resolve K2, an antecedent must be substituted to &quot;he&quot;; this antecedent will be Berg (because argtunents in OPFJtA are typed).</Paragraph> <Paragraph position="9"> The DP, S of situation becomes :</Paragraph> </Section> <Section position="5" start_page="25" end_page="27" type="metho"> <SectionTitle> 4.2 TREAI\[MENT OF ELJ~IPSIS </SectionTitle> <Paragraph position="0"> One of the most con~aon \]phenomena of dialogue is ellipsis such as in : and a symphony The interpretation of (b) is &quot;Britten a cQuposd Loulou&quot; (i.e. VP-ellipsis); and (c) is to be interpreted as &quot;Berg a compos6 tme symphonie&quot;. The ad hoc treatment proposed in \[SEDOGBO, 85\] fails in 1lmny cases; as we mentioned, since the logical formulas in OPERA are equivalent to first-order logic formulas.</Paragraph> <Paragraph position="1"> An interesting extension to VP-ellipsis in DR-theory is described by KLEIN (cf \[KLEIN 84\])deg KLEIN introduced the notion of predicate-DRS and pro- null posed an indexation of NP predicate-DRS and VP predicate-DRS in a DRS.</Paragraph> <Paragraph position="2"> We will not propose here how this extended DRS can be implemented; but will exploit the parallelism between our logical formulas and DRSs.</Paragraph> <Paragraph position="3"> Each sentence will be translated into two partiallogical formulas (noted PLF) of the form <x,f>, where x is a variable or an individual and f a logical formula. null We assume that the composition of PLF(NP) and PLF(VP) gives the translation of the sentence.</Paragraph> <Paragraph position="4"> The first PLF is implicitly indexed by the NP and the second PLF is indexed by the VP.</Paragraph> <Paragraph position="5"> The DRS of current situation must be modified in the following way : KSS(n)= \[USS,CONSS\] with USS defined as above, and CONSS containig PLF(VP). Given a VP-ellipsis s, it will be translated into the the PLF: <i,fl>. The DRS of current situation contains in its CONSS a PLF :<j,f2>.</Paragraph> <Paragraph position="6"> The VP-ellipsis is treated in the following way: i) ~lification of i and j 2) the c~upositi~ of <i,fl> and <j,f2> produces the translation t 3) t is translated into a DRS K and KSS is built as described in session (4.1) We will illustrate the VP-ellipsis treatment by processing the dialogue above: (a) is translated into the formula compose(Berg,Loulou) and PLF(NP) = <Berg,Berg>, PLF(VP)= <i,cempose(i,Loulou)>.</Paragraph> <Paragraph position="7"> The translation t of (a) is then obtained by ccraposition of PLF(NP) and PLF(VP); thus t: compose(Berg,loulou).</Paragraph> <Paragraph position="8"> The evaluation of t will augment the dialogue situation of a new DRS of situation containing PLF(VP). The translation of (b) will produce a PLF(NP)= <Britten,Britten> and a PLF(VP)= <i,p>.</Paragraph> <Paragraph position="9"> <i,p> will then be unified to the PLF(VP) contained in the CONSS of KSS , i.e <i,p> is unified with <i,ccmpose(i,Loulou)>; the composition of the two PLF <Britten,Britten>, <Britten,canpose(Britten,Loulou)> will produce the logical formula &quot;ccmpose(Britten,Loulou)&quot;.</Paragraph> </Section> class="xml-element"></Paper>