<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2189">
  <Title>ROBUST METHOD OF PRONOUN RESOLUTION USING FULL-TEXT INFORMATION</Title>
  <Section position="3" start_page="0" end_page="7158" type="metho">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Resolving t&gt;ronoun reference is a diifieult task that requires consideration of both linguistic and cognitive aspects of a language. As a linguistic phenomenon, the use of pronouns is treated a.s a co-referential prohlem in which both the antecedent and the pronoun co-refer to some object. From this point of view, finding the object that is co-referred to by a pronoun is the main problem in t)ronoun resolution, and much research has been devoted to focusing on or inferring the referent object by considering the grammatical and semantic roles of each entity in the sentences \[Sidner, 1983; Brennan, 1987; Kameyama, 1993\]. This task is especially difficult when the referent object is not explicitly stated in a text, and common sense and deep inference are required in order to figure out the object, sLs in the classic problems descrit)ed by Charniak \[Charniak, 1973\]. Since this approach of considering the grammatical and semantic roles of each entity depends heavily on accurate syntactic anMysis avd semantic analysis, it is not yet applicable to practicM systems.</Paragraph>
    <Paragraph position="1"> However, if we do not aim for perfect analysis, a simple syntactic-based heuristic rule for selecting a correct antecedent from several candidate noun phra~ses i)erforms quite well, especiMly in technical documents such as computer manuals, in which we can usuMly expect an explMt antecedent within the same sentence or in a previous sentence. In this domain, a correct ante(:edent can be selected in ahnost 90% of all causes without any world knowledge other than simple semantic constraints \[Hobbs, 1978; Walker, 1989; Lappin, 1990\].</Paragraph>
    <Paragraph position="2"> Moreover, several heuristic rules can be combined to improve the accuracy of the analysis \[Rich, 1988; Carboi,ell, 1988\].</Paragraph>
    <Paragraph position="3"> This approach of resolving pronoun reference by applying simple heuristic rules seeins to be ade.quate for a practical naturM language processing system, yet in order to achieve a success ratio of over 90%, some kind of knowledge processing is required, such as the use of world knowledge or deep inference mechanisms for constructing and referring to a discourse structure. While the advantage of knowledge processing is widely recognized, this approach presupposes a large quantity of knowledge resources, and leads to a knowledge acquisition bottleneck. In order to solve this problem, wmous studies have been done on methods of using on-line text databa~ses with less human intervention for word sense disambiguation attd structural disambiguation \[Jensen, 1987; Nagao, 1990; Uramoto, 1991; Hindle, 1993\]. These methods can be applied to knowledge processing in pronoun resohltion; however, no research has yet reveMed sutficient world knowledge to cover general 1)roblems. In other words, methods of using world knowledge have not reached a level sufficiently mature for them to be used in broad-coverage systems.</Paragraph>
    <Paragraph position="4"> This paper prol)oses a simple and robust approach that utilizes inte.r-sentential information, extracted from a source text by means of a simple algorithm, to improve the accuracy of pronoun resolution. For example, (:olloeation patterns within a text offer information that eorrcsl)onds to ease frames in world knowledge, and word frequency also gives information reh!wmt to the tot)it or focus of the subjects. Thus, instead of using outside knowledge resources, such information serves as world knowledge appropriate to the narrow domain of the source text. The effectiveness of each type of information extracted from a source text is evaluated in the light of the results of experiments on comlmter lnanuals.</Paragraph>
    <Paragraph position="5"> In the next section, we introduce three effective factors in the selection of an antecedent from candidate noun phra.ses. Then, in the third section, we  specify the implementation of this method. Finally, in the fourth section, we evaluate the effectiveness of this approach on the basis of the results of an experiment. null 2 Three factors for evaluating salience in candidates In our approach, pronoun resolution basically consists of collecting candidate noun phrases and selecting the most preferable candidate as the antecedent of a pronoun by applying several rules to filter out inappropriate candidates and to attach preferences to appropriate candidates. Rules are divided into two types. One type represents grammatical constraints that must be satisfied, such as number and gender agreement. Since rules of this type can filter out inappropriate candidates, we apply them at an early stage of pronoun resolution. The remaining rules constitute the other type, which attaches a preference to each candidate noun phrase. After inappropriate candidates have been filtered out by the former rules, the latter rules determine the most appropriate candidate by measuring the sMience of each remaining candidate noun phrase. Thus, the latter rules are important for selecting the exact antecedent from the remaining candidates and for improving the accuracy of pronoun resolution.</Paragraph>
    <Paragraph position="6"> In this section~ we describe three effective factors that utilize full-text information for measuring the salience of each candidate noun phrase. The reasons for their effectiveness are that they cover many aspects of linguistic phenomena and that their interpretation is simple enough to be used in a practical system.</Paragraph>
    <Section position="1" start_page="1157" end_page="7158" type="sub_section">
      <SectionTitle>
2.1 Collocation patterns within a
</SectionTitle>
      <Paragraph position="0"> source text In previous approaches, semantic constraints have been among the most basic factors for filtering out candidates that would be inappropriate as modifiers of the modifiee of a pronoun. However, in order to apply semantic constraints with broad coverage, a large amount of knowledge is required. For example, in processing a sample sentence provided by Hobbs \[Hobbs, 1978\], The castle in Camelot remained the residence o/ the king until 536 when he moved it to London, the following knowledge must be supplied in order to filter out the candidates 536, castle, and Camelot, and leave the correct antecedent, residence:  * Dates cannot move.</Paragraph>
      <Paragraph position="1"> * Places cannot l~love.</Paragraph>
      <Paragraph position="2"> * Large fixed objects cannot move.</Paragraph>
      <Paragraph position="3">  In order to apply this knowledge, we also prcsul)pose a correct analysis that categorizes each nmm phrase as a date, place, large fixed object, and so on. Since many of the words in these noun phrases have word sense ambiguities, it is not practical to presuppose the correct application of such knowledge. Assembling a large body of knowledge poses another major problem.</Paragraph>
      <Paragraph position="4"> instead of such worhl knowledge, collocation patterns (namely, modifiee-modifier relationships) extracted from a discourse can be applied. Since word sense is usually unified within a discourse, and most words with the same lemma are frequently repeated \[Gale, 1992; Nasukawa, 1993\], the collocation patterns in the same discourse provide valuable data for determining whether a candidate can modify the modifiee of Ct pronoun. For example, if the sentence He moved his residence is found in the discourse, this information indicates that the word residence can be the object of the verb move. Thus~ the inh)rmation works as a selectional constraint that the candidate can be an argument of a predicate that dominates the pronoun.</Paragraph>
      <Paragraph position="5"> Moreover, since statements tend to be repeated in a discourse, the existence of an identical collocation pattern in a discourse may support selection of the candidate as the antecedent. In this sense, the preference for an identicM collocation pattern also reflects case role persistence preference and syntactic parM1elism preference, prolmsed by Carbonell and Brown \[Carbonell, 1988\]. The case role persistence rule prefers a candidate noun phrase that filled an identical case role in an earlier sentence. For example, after the sentence Mary gave an apple to Susan, Susan is referred to by her in John also gave her an apple, while Mary is referred to by she in She also gave John an apple.</Paragraph>
      <Paragraph position="6"> The syntactic 1)arM1elism rule prefers a candidate noun phrase that preserves its surface syntactic role from the first of two or more coordinate clauses. For example, in  The girl scout leader paired Mary with. Susan, but she had paired her with Nancy last time, she refl,'rs to leader, and her refers to Mary, whereas in The girl scout leader paired Mary with Susan, but she had paired Nancy with her last time, she refers go leader, ait(\[ her refers to ,S'usan. By referring to the i(lenticM collocation Iiatterns, we (:~ut resolve all the pronouns in the above examples (:orrectly. null Since the idcntitication of modifier-modifiee rela tionshii)s is a basic feature of syntactic analysis, a procedure for identifying identical collocation patterns is not a hard task, as long as M1 of the sentences are parsed by a single system and share a single l'orm~flism for expressing modifier-modiiiee relationships.</Paragraph>
    </Section>
    <Section position="2" start_page="7158" end_page="7158" type="sub_section">
      <SectionTitle>
2.2 l~'equency of repetition in preced-
</SectionTitle>
      <Paragraph position="0"> ing sentences A characteristic of the i)ronolnimflization on whi(:h the centering apln'oach \[Sidner, 1983; Ilr(,mlan, 1987; Kameyama, 1993\] is l);Lsed is that an object in focus is likely to hc pronomimtlized. If this characteristic is expanded to all definite anaphoras, which include detinite noun phrases as well as pronouns, a candidate noun phrase that is in focus inay be repeated as a definite noun phrase before it is pronomiilalized. Thus, the frequen(:y in 1)receding sentences of a nouu i)hrase with the same lemma ~s ~t candidate noun l)hr~use (:aa be an index for the preference with whi(:h it is selected as the antecedent. The 1)roeess for a.ssigning this 1)reference consists of a simple string match that cheeks words with the same lcmma in preceding sentences.</Paragraph>
      <Paragraph position="1"> In addition, when the source text is marked up with SGML or other su(-h tags, the roles of some phrases such as titles mid headings call he easily re(&gt; ognized, and words with such roles tend to represent the topics of the sentences fifth)wing them. Thus, ~(\[ditional preference can be assigned by checking the tags of each word.</Paragraph>
    </Section>
    <Section position="3" start_page="7158" end_page="7158" type="sub_section">
      <SectionTitle>
2.3 Syntactic position
</SectionTitle>
      <Paragraph position="0"> As shown by tIobtis \[ltobl)s, 1978\], a heuristic t'ule favoring subjects over objects l)erforlns remarkably well in English text. By traversing th(' surface parse tree of a sentence, a 1)referen(:e vMue can be provided for each candidate noun phrase according to its syntactic position. This factor has mt advantage over other fa(&gt; tors shown in 1)revious subsections in the sense that it: assigns a definite ranking for ea(:h candidate noun I)hrase, since ea(:h o(:eul)ies a syntactic position in a text. Thus, this factor l)rovides a default value for the l)reference of ea(;h (:an(lidate itOUlt t)hrase when no oth('r factor provides wfiid information, and it is ad('(luat(~ for a r(It)ust approach since it is basically assigned t)y traversing tho surface l)arse trees of a sen t.enee.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="7158" end_page="7158" type="metho">
    <SectionTitle>
3 Implementation
</SectionTitle>
    <Paragraph position="0"> In this sect;ion, we describe the actual implelnelltation of (;he l)rottoUll resohltiott procedllre ill ml Fmglish-to-Japanese. machine translation system, Shalt2 \[Takeda, 1992\].</Paragraph>
    <Paragraph position="1"> The procedure consists of two steps:  1. Extraction of ca.ndidat(:s for an mltec(,dent 2. Selecti(m of I;he correct antece(tent fl'om the candidates.</Paragraph>
    <Paragraph position="2"> \[11 or(let t(/ achieve ttigh(!r ac(;llracy itl l)rolloull resolution with robust lirocessing, our straeegy consists of 1. Extending a list of ('andidate noun ptm~scs so tht~t it does noL ex(:lude a correct antecedent 2. \]leferring t() all information in the source text that can be syntactically extracted without referring to outside knowledge resources, in orde, r to sele(:t the corre('.t ante(:e(lent.</Paragraph>
    <Section position="1" start_page="7158" end_page="7158" type="sub_section">
      <SectionTitle>
3.1 Extraction of candidates
</SectionTitle>
      <Paragraph position="0"> 'fo ensure that the correct antecedent is in(:luded in a list of candidate noun t)hrases, candidates are extra(:ted from exa(:tly two sentences with lninimum tiltering. First, the system cheeks whether any noun tihrases earlier in the same sentence satisfy the num-ber and gender constraints. It then checks the t)reced ing sentences in order of proximity until candidates have bct}ll YOlllI(\[ ill exactly two sentell(:es.</Paragraph>
      <Paragraph position="1"> l)uring the extraction of the candidates, the system filters out noun phrases that do not satisfy the numl)er and gender constrmnts, and Mso direct modifiees of the pronoun and its arguments, so that a noll-reIlexive 1)rOllOlllt all(t its anl;eced(mt nlay not occur in I:he same siml)lex sentence, as would be the case if data were the antecedent of it in the following sentenc(~; The device that writes onto a magnetic di.sk and reads data from it is called a disk drive.</Paragraph>
    </Section>
    <Section position="2" start_page="7158" end_page="7158" type="sub_section">
      <SectionTitle>
3.2 Selection of an antecedent
</SectionTitle>
      <Paragraph position="0"> In our implementation, the preference values provided by the algorithms described in the following paragraphs are combined into a single value, and the candidate noun phrase with the largest preference value is selected as the antecedent.</Paragraph>
      <Paragraph position="1"> Preference according to the existence of identical collocation patterns in the text As a preference value that indicates the satisfaction of selectional constraints and repetition of an identical statement, we assigned a constant value 3 for a candidate that has an identical collocation pattern with the modifiee of a pronoun within the source text. Furthermore, in order to extend the use of collocation patterns as knowledge on seleetional constraints, an on-line synonym dictionary \[3\] is referred to, and thus a collocation pattern with a synonym can support candidates other than exactly identical collocation patterns.</Paragraph>
      <Paragraph position="2"> Preference according to the frequency of repetition in preceding sentences In order to provide a larger preference value for closer and more frequent occurrences of a lemma, the preference value is given by the total score calculated according to the following formula, for each appearance of a noun phrase with the same lemma as tim candidate noun phrase that is found within the ten preceding sentences:  (Number of sentencea to the identical noun phrase)+1. Preference according to syntactic position Among the candidate noun phrases, a candidate in a closer sentence, or the one nearest the beginning of the same sentence is preferred. Besides the left-to-right order within a sentence, a negative preference value is given for tile distance (number of sentences) from the sentence that contains the pronoun to the sentence that contains the candidate. While the order of preference of candidates that is obtained in this manner is similar to that given by the naive algorithm proposed by Hobbs \[Hobbs, 1978\], our algorithm is much simpler, and does not even require the results of syntactic analysis.</Paragraph>
    </Section>
    <Section position="3" start_page="7158" end_page="7158" type="sub_section">
      <SectionTitle>
3.3 Example
</SectionTitle>
      <Paragraph position="0"> Figure 1 gives an example of system output that contains data on the preference of each candidate antecedent for a pronoun in a sample text extracted f,'om the second chapter of a computer manual \[2\] in the mam, er described in the previous paragraphs.</Paragraph>
      <Paragraph position="1"> In this figure, the number in brackets before each sentence indicates the sentence number in the text.</Paragraph>
      <Paragraph position="2"> As shown by these numbers, the output consists of eleven consecutive sentences, front the 104th to the 114th in the second chapter of the manuM.* The order of candidates following the message Candidates for the referent of CFRAME106579 (&amp;quot;it&amp;quot;) are : reflects tile order of preference values obtained by referring to the positimt of each candidate. As shown in this list, key is the most preferable candidate from the viewpoint of syntactic position. In this candidate list, CFRhMEwuwo indicates an instance of each content word in the dis(:ourse. Information on the position and on tile whole sentence can be extracted from each of these CFlt.AMEs. A number in arrowhead brackets next to CFRAMEu~uu~u, such as &lt;ll3&gt;, indicates the number of the sentence in which it occurs. A number in parentheses, such as 0.48571432 in key (0.48871432), indicates the preference value obtained by referring to the frequency of repetition. 2 Thus, from the viewpoint of the number of repetitions, cursor is the most preferable candidate for the aI, tecedent. At the bottmn of this figure, information on modifier-modifiee relationships that support candidates is shown. In this case, there is a collocation pattern such that cursor modifies the verb reaches, which is the modifiee of tile pronoun it; tiros, this information prefers cursor for the antecedent of it.</Paragraph>
      <Paragraph position="3"> Finally, after combination of all the preferences, cursor is selected as the most preferable antecedent of it.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>