<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2184">
  <Title>EXTENDING DRT WITH A FOCUSING MECHANISM FOR PRONOMINAl. ANAPHORA AND ELLIPSIS RESOLUTION</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1. INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> Pronominal anaphora resolution, as part o1' a more general process of anaphora resolution, is a determinant step in constructing a semantic representation of a text.</Paragraph>
    <Paragraph position="1"> Although &amp;quot;general cognitive processes DO play a role in establishing anaphoric dependencies (...)&amp;quot; (Kempson, 1990 p.14), inference is, in computational terms, a very expensive process, both for the amount of processing involved and for the extension of the knowledge bases required. Therefore, any system aiming at efficiency in anaphora resolution should minimize the role of inference.</Paragraph>
    <Paragraph position="2"> As far as DRT is concerned, the construction rule for pronouns slates that the referent introduced by the pronoun should be bound to a suitable referent, chosen among those that are accessible (Kamp and Reyle, 1993 p.122). The accessibility is based on semantic constraints and is expressed by the structure of DRS representing the text. However the suitability of referents is ill-defined.</Paragraph>
    <Paragraph position="3"> Another perspective for anaphora resolution is founded on the principle of relevance, i.e. on &amp;quot;the presumption that every utterance is selected to convey the intended interpretation while imposing to the hearer the least amount of processing effort in constructing that interpretation&amp;quot; (Kempson 1990 p.17). Focusing/ centering theories (Grosz; Sidner; Brenn,'m, Friedman and Pollard et al.) can be considered as having this perspective. They try to keep track of the focus of attention along the text and bind pronouns preferentially to focused entities. The choice of antecedents is based on pragmatic constraints, which put an ordering on preferences between ,antecedent czmdidates.</Paragraph>
    <Paragraph position="4"> Cormack proposes the integration of focusing and DRT, &amp;quot;(...) adding semantic constraints to a model of attention in discourse&amp;quot; (Cormack, 1992 p.5). This integration compensates for two shortcomings of DRT: it considers too many possibilities for anaphoric binding and doesn't provide an ordering between antecedent candidates. From the focusing point of view, the addition of semantic constraints, provided by DRT, to the pragmatic ordering further restricts the determination of possible antecedents.</Paragraph>
    <Paragraph position="5"> We analyzed Cormack's proposal, and found out that it was lacking some features that we consider more adequate, as it will be shown in the next few sections.</Paragraph>
    <Paragraph position="6"> Therefore we adapted it, and applied the modified version to the processing of texts written in Portuguese. The scope of d~osc mcthods was widened to cover sentences containing restrictive relative clauses and subject ellipsis. Tests were conceived and applied to probe the adequacy of proposed modifications when dealing with processing of current texts.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="1129" type="metho">
    <SectionTitle>
2. SIMPLE SENTENCES
</SectionTitle>
    <Paragraph position="0"> 2.1. Alterations to DRT Cor,nack defends that pronouns of the current sentence can only have access to two groups of referents: focused referents and those unfocused ones that were introduced by the preceding sentence. Referents not fitting any of these two groups can be forgotten. Let us look at an example (Connack, 1992 p.350): (la) John took apart the chest 0 f drawers.</Paragraph>
    <Paragraph position="1"> (lb) it was full of clothes pegs.</Paragraph>
    <Paragraph position="2"> The DRS representing the first sentence will be (foct, sed referents are shown on the left, unfocused ones on the right):</Paragraph>
    <Paragraph position="4"> The second sentence introduces another DRS. Anaphors are resolved with referents of previous DRS,</Paragraph>
    <Paragraph position="6"> and then previous DRS can be &amp;quot;R~rgotten&amp;quot;:</Paragraph>
    <Paragraph position="8"> Referent John, who was introduced by (la), was only available for anaphor resolution in (lb). Since it was never focused, it is &amp;quot;forgotten&amp;quot;. This means that it is no longer included in the referents of the DRS representing the text after processing of the second sentence, becoming unavailable as antecedent candidate for pronouns in following sentences. This claim may seem a little strange if we look at (lc) as an acceptable third scntenee: (lc) lie didn't like dmir color.</Paragraph>
    <Paragraph position="9"> Two other aspects of Cormack's representation led us to prefer to keep to the original DRT formalism. First, Cormack's representation is too conditioned by pronominal anaphora resolution. Referents that become unavailable for pronominal reference, and are therefore &amp;quot;forgotten&amp;quot;, may still be cospecified hy definite descriptions. Eliminating them from tbe representation would be a limit to the possibilities of expanding the system in the fltture. Second, &amp;quot;forgetting&amp;quot; conditions introduced by previous sentences leads to a situation where the DRS representing Ihe text at a given moment will contain little information about the text, and no information at all about some of the &amp;quot;surviving&amp;quot; referents* For instance, looking at the last DRS presented, we no longer know what entity introduced referent cl.</Paragraph>
    <Paragraph position="10"> 2.2. Focusing algorithms Most focusing theories keel) referents that can be relevant in future anaphora resolution in focus stores. Sidner considers two groups of focus stores, which in a very short and simplistic way can be described as: those related to agent tAG) role: actor locus (AF) - AG of current sentence or previous AF, if current sentence has no AG; potential actor focus list (PAFL) - other animate referents of current sentence; actor focus stack (AFS) - previous AFs; those related to other thematic roles: discourse focus (DF) - null text knowledge base where all the restrictions upon referents are present.</Paragraph>
    <Paragraph position="11"> 2 sue (Sidner. 1979), (Cormack, 1992) for details about this ranking potential discourse locus list (PDFL) - referents of current sentence excluding DF; discourse focus stack (DFS) - previous DFs.</Paragraph>
    <Paragraph position="12"> In determining the antecedent of a pronoun, algorithms go through some preliminary considerations (such as recency rule) and a basic ordering of focus stores.</Paragraph>
    <Paragraph position="14"> Although taking Sidner's algorithms as a starting point, Cormack renounces the distinction between actor focus and discourse focus, in the final part of her work.</Paragraph>
    <Paragraph position="15"> The algorithms become more simple but they loose in discriminatory power. This is particularly more significant in a language like Portuguese, where nominals can only be masculine or feminine (not neuter)* In a text like (2a) O Jofio escreveu um livro.</Paragraph>
    <Paragraph position="16"> John wrote a book. (AF = John, DF = a book) (2b) A Maria lets-().</Paragraph>
    <Paragraph position="17"> Mary read it.</Paragraph>
    <Paragraph position="18"> eliminating the distinction between AF and DF would lead to Jodo (lohn) being proposed as preferred antecedent of the masculine pronoun o (it). Rejecting this binding would require an appeal to inference, wlfich is something that we want to minimize. Keeping AF - DF distinction will also be significant in dealing with another phenomenon very common in Portuguese: subject (SU) ellipsis.</Paragraph>
    <Paragraph position="19"> Recency rule &amp;quot;if the pronoun under consideration occurs in the subjcct position, and there is an alternate focus list noun phrase which occurs as the last constituent in the previous sentence, test that alternate focus list phrase for co-st)ecification before testing the current focus. (...)&amp;quot; (Sidner, 1979 p.144).</Paragraph>
    <Paragraph position="20"> Sidner admits that &amp;quot;the recency rule makes focussing seem somewhat ad hoc&amp;quot; (ibid.), Carter states that &amp;quot;its inclusion in SPAR led to considerable inaccuracy&amp;quot; (Cartes&amp;quot; 1987 p.114) and Cormack decides to ignore it too (Cormack, 1992 p.54), ttowever, it seems that, in Portuguese, this rule should be considered for pronouns in AG position: (3a) A Maria i deu um livm a Anaj.</Paragraph>
    <Paragraph position="21"> Mary i gave Annj a book.</Paragraph>
    <Paragraph position="22"> If tile agent of the ncxt sentence is Mary there are two ix)ssit)ilities of pronominalization: the prontmn ela (she) or the null pronoun (~ (SU ellipsis). This last option will be l)mferrexl: (3b) ~i comprara-o num leilfio.</Paragraph>
    <Paragraph position="23"> C/i had bought it at an auction*  But if the agent of the next sentence is Ann, the only possibility of pronominalization will be the cxplicit pronoun ela (she): (3b') Elaj Icu-o.</Paragraph>
    <Paragraph position="24"> Shej read it.</Paragraph>
    <Paragraph position="25"> So thc speaker will tend to use a null pronoun in AG position to eospccify the agcnt of the previous sentcncc, reserving the explicit pronoun a use that conforms with thc recency rule.</Paragraph>
    <Paragraph position="26"> Intrasentential anaphora Carter inserts intrasentential candidates (ISC) between current feel and potential loci, in the basic ordering. Cormack distinguishes between focused ISC and remainder oflSC. In our implementation this distinction seemed unnecessary and we decided to insert ISC alter potential foci, in thc basic ordering. A special casc of ISC is the reflexivc pmn(mn se (himself/herse!f/itself/ themselves). We always bind it to the agent of the scntcnce.</Paragraph>
    <Paragraph position="27"> (4) O camelo i dcitou-se i na arcia.</Paragraph>
    <Paragraph position="28"> The camel i laid (itse!f i) down on the sand.</Paragraph>
    <Paragraph position="29"> Intrasentential catapbora In our implementation, syntatic parsing is done according to grammar (levclopment formalisms hased on barricrs, movcment and binding (Lopcs 1991). It is an cxtension of thc extraposition grammar formalism (Pereira 1981) and allows for movement of constituents of a scntencc in a rcstrictcd area delimited by harricrs. The resulting synUltic trcc will always show the intcrnal arguments of the verb on it's right, no matter what positions they had in the original sentence. For instance, the syntatic trec for  (5) Near her, the blond girl saw a man.</Paragraph>
    <Paragraph position="30"> will be:</Paragraph>
    <Paragraph position="32"> lhe blond girl saw a man near her The anaphora resolution process works on the restdts of the syntatic parser, so this kind of cataphora will be trc~lted as intrasentential anaphora.</Paragraph>
    <Paragraph position="33"> Subject ellipsis As mentioned above, this is a very common phcnomcnon in Portugncse language. Null pronoml in AG position seems to behave differently from onc in non-AG position. In thc first case it cospecifies AF or a combination of foci including AF: (6a) A Maria i dccidiu ofcrcccr aquele perfume h Arm. Mary i decided to offer Ann that perfume. AF = Mary (6b) ~i gostava muito dole.</Paragraph>
    <Paragraph position="34"> (b i liked it very much.</Paragraph>
    <Paragraph position="35"> A null pronoun in non-AO position cospccifics DF or a combination of foci including DF:  syntactic agreement and consistency with world knowlextge to a ratification procedure, to be appliexl after completion of focusing process. Efficiency can be improved if inexpensive number and gcnder agreement and reflexivity verificatk)ns arc included in the focusing proccss. Thus, scvcral inadcqrmte candidates can be ruled out without a call to the ratification procedure.</Paragraph>
  </Section>
  <Section position="5" start_page="1129" end_page="1131" type="metho">
    <SectionTitle>
3. SENTENCES CONTAINING
RESTRICTIVE RELATIVE CLAUSES
</SectionTitle>
    <Paragraph position="0"> Going beyond simple sentcnces, we widcncd thc scope of the prescntcd methods to includc sentcnccs with restrictive relative clauses (for short, we'll just use the form relative clauses in the remainder of this paper).</Paragraph>
    <Paragraph position="1"> Rules for focus movement and refcrcnts accessibility were formulated and tests werc dcsignetl to probe their adcqt, acy. In this secti(m we refer to the results of a qucstionnairc answered by 40 collcge students.</Paragraph>
    <Paragraph position="2"> Focus movement (8a) O Joao leu unl livro i.</Paragraph>
    <Paragraph position="3"> John read a book i. DF = a book (8b) O homemj que o i cscreveu morreu.</Paragraph>
    <Paragraph position="4"> The manj who wrote it i died.</Paragraph>
    <Paragraph position="5"> (8c) Os eruditos cnalteceram-no i v j ? muito.</Paragraph>
    <Paragraph position="6"> Erudite people praised him~it i v j ? much.</Paragraph>
    <Paragraph position="7"> According to focusing rules, tile pronoun in (8c) cospecifies DF of (8b). ff lnXmouns in relative clauses were able to influence focus then (81)) would confirm a book as DF and this would be the antecedent of the pronoun in (8c). That doesn't sccm to be the case. The intuitively preferred antecedent is the man. Examples like this show that pronouns occurring within relative clauses dofft seem to inlluence focus movement. This colmlusion was confirmed by 83% of the answers to the alx)ve mentioned questionnaire.</Paragraph>
    <Paragraph position="8"> Access of following sentences to relative clause referents Referents introduced 1)y the relative clause arc acccssiblc hut arc not preferred to main clause rcfcrcnts. The qucstionnairc prcscnted the text:  (9) 0 homem a quem u,n ladr~o roubou o rcldgio chamou a polfcia. Ele ...</Paragraph>
    <Paragraph position="9"> The man whom a thief stole the watch from calh:d the police, lie ...</Paragraph>
    <Paragraph position="10"> 58% of the continuations proposed bind the pronoun to the main clause referent the man while only 28% indicate binding with file relative clause referent a thief. Access of the relative clause to main clause referents (10) O Joao deu um livro i at) ahmo qnc o i merecia. John gave a book i to the student who deserved it i. Pronouns in the relative clause can cospecify both main clause referents or focus stores. The first situation seems to be preferred except, perhaps, for pronouns in AG position, that show a weak preference (suplx)rted by 61% of the answers) for cospecification with AF or a member of PAFL.</Paragraph>
    <Paragraph position="11"> Access of the main clause to relative clause referents (11) O homem que escreveu um livro i deu-o i a Maria. The man who wrote a book i gave it i to Mary.</Paragraph>
    <Paragraph position="12"> Pronouns in the main clause, occunilLg after the relative clause, can cospecify it's rclerents, lake Cormack, we conskler access to focus storcs to be more likely, lint this preference was not confirmed by the results of the questionnaire (60% of the answers were against).</Paragraph>
    <Paragraph position="13"> Access of relative clause to relative clause (12) O homem qtte a Maria i viu escreven um livro title a i imprcssionou.</Paragraph>
    <Paragraph position="14"> The man who was seen by Mary i wrote a book that impressed her i.</Paragraph>
    <Paragraph position="15"> Pronouns in the second relative clause can cospccify referents of the first one. ttowever, it seems that main clause referents should be prelerrtxl as antecedents. The example used to test this preference was not very clear and so we've got 63% of negative answers.</Paragraph>
    <Paragraph position="16"> Transitive access to a main clause (13) A Maria i casou corn o cliente qlce conlprou 0 livro que ela i e~reven.</Paragraph>
    <Paragraph position="17"> Mary i married the client who bought the book that she i wrote.</Paragraph>
    <Paragraph position="18"> Pronouns in ,'1 nested relative clause cnt |cospccify main clause referents. Preference seems to be given to antcccxlent candidates of the main clause over those el: the nesting relative chmse, but this hypothesis was not testexl.</Paragraph>
    <Paragraph position="19"> Transitive access to a relative clause (14) O cliente que comprou () livro que a emprcgada i escrcveu casou corn ela i.</Paragraph>
    <Paragraph position="20"> The client who bought the book that was written by the employee i married her i.</Paragraph>
    <Paragraph position="21"> Pronouns in the main clause can cospccify nested relative clntnse referents. Candidate nnteccxlents occurring in the nesting relative clause seem to be preferred though. This preference is supported by 75% of the answers.</Paragraph>
    <Paragraph position="22"> Ordering autecedent candidates We can summarize this analysis in the following rules lot predicting antecedents. These rules were implemented without significant changes to the algorithm establishcd lbr simple sentences.</Paragraph>
    <Paragraph position="23"> Relative clause pronotms: AG position: : main clause AG not null: Ab', PAFL, main clause refs., remainder of focus stores non-AG position: main clause refs., lbcus stores ,Main clause pronouns: Prcc(xling a relative clause: focus stores \[q)llowing a l+clative clause: idem excluding stacks, relative clause rel~., slacks Folk)wing sentence pronouns-main clause refs., relative clause refs., slacks Nested relative clauses: They have transitive access to main clause refs. Main clause pronouns prefer nesting clause refs. to nested clause ones.</Paragraph>
    <Paragraph position="24"> Relative clauses as conditionals Both Kamp (1993 p.81) and Cormack (1992 p.347) propose a &amp;quot;flat&amp;quot; treatment of relative clauses. Both it's referents (with the possible exception of proper names) and conditions are introduced in current DRS.</Paragraph>
    <Paragraph position="25">  adequate to all kinds of relative clauses in Portuguese, namely those whose verb is in subjunctive mood. (17) Um agricultor que tenha um burr() bate-lhe. A farmer who (subjunctive of ~) a donkey beats it.</Paragraph>
    <Paragraph position="26"> This kind of sentences is associated to non-factual, hypothetical presuppositions and is semantically equivalent to an implication relation between two clauses: (17') Seum agricultor tern um burro entfio bate-lhe. If a farmer owns a donkey then he beats it.</Paragraph>
    <Paragraph position="27"> So, our implementation represents this kind of sentences as conditionals:  Our rules for anaphora resolution will then be applied as usual, taking in consideration both focusing and semantic (DRT-detennined) acessihility constraints.</Paragraph>
  </Section>
  <Section position="6" start_page="1131" end_page="1131" type="metho">
    <SectionTitle>
4. TESTING
</SectionTitle>
    <Paragraph position="0"> Tests were conceived with the only purpose of probing the adequacy of proposed modifications. One of the tests, the questionnaire, has already been inentioned. It consisted of two parts. In the first one there were short texts (2-4 sentences) where some referents were introduced. The last sentence was always incomplete and contained a pronoun. The continuation proposed by the student was supposed to show which co-specification he had chosen. Since the evaluation of this part might be influenced by intuition, it was committed to 3 independent evaluators, who were found to agree on 80% of the answers. The second part consisted of texts of the same kind, but where all sentences were co,nplete. The student was asked to identify explicitly the co-specification of a pronoun introduced by the last sentence. The results concerning relative clauses were presented in last section. Recency rule and rules for subject ellipsis were confirmed respectively by 77% and 85% of the answers.</Paragraph>
    <Paragraph position="1"> The two other tests consisted of applying the rules for relative clauses to all anaphome found in current texts, and whose antecedent or anapho,&amp;quot; were introduced by a relative clause. &amp;quot;Fhe first target text was a novel by a famous Portuguese writer of lhe end of last century, Eqa de Queiroz (19(/0). The news of a Portuguese news agency (Lusa, 1993) provided 637 kbytes of fresh (June93) raw material for the last test. The rules performed correctly in respectively 96% and 92% of the cases.</Paragraph>
  </Section>
class="xml-element"></Paper>