<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-1011">
  <Title>Evaluating a Focus-Based Approach to Anaphora Resolution*</Title>
  <Section position="4" start_page="0" end_page="74" type="metho">
    <SectionTitle>
2 Focus in Anaphora Resolution
</SectionTitle>
    <Paragraph position="0"> The term focus, along with its many relations such as theme, topic, center, etc., reflects an intuitive notion that utterances in discourse are usually 'about' something. This notion has been put to use in accounts of numerous linguistic phenomena, but it has rarely been given a firm enough definition to allow its use to be evaluated. For anaphora resolution, however, stemming from Sidner's work, focus has been given an algorithmic definition and a set of rules for its application. Sidner's approach is based on the claim that anaphora generally refer to the current discourse focus, and so modelling changes in focus through a discourse will allow the identification of antecedents.</Paragraph>
    <Paragraph position="1"> The algorithm makes use of several focus registers to represent the current state of a discourse: CF, the current focus; AFL, the alternate focus list, containing other candidate foci; and FS, the focus stack. A parallel structure to the CF, AF the actor focus, is also set to deal with agentive pronouns. The algorithm updates these registers after each sentence, confirming or rejecting the current focus. A set of Interpretation Rules (IRs) applies whenever an anaphor is encountered, proposing potential antecedents from the registers, from which one is chosen using other criteria: syntactic, semantic, inferential, etc.</Paragraph>
    <Section position="1" start_page="74" end_page="74" type="sub_section">
      <SectionTitle>
2.1 Evaluating Focus-Based Approaches
</SectionTitle>
      <Paragraph position="0"> Sidner's algorithmic account, although not exhaustively specified, has lead to the implementation of focus-based approaches to anaphora resolution in several systems, e.g. PIE (Lin, 1995). However, evaluation of the approach has mainly consisted of manual analyses of small sets of problematic cases mentioned in the literature. Precise evaluation over sizable corpora of real-world texts has only recently become possible, through the resources provided as part of the MUC evaluations.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="74" end_page="74" type="metho">
    <SectionTitle>
3 Coreference in LaSIE
</SectionTitle>
    <Paragraph position="0"> The LaSIE system (Gaizauskas et al., 1995) and (Humphreys et al., 1998), has been designed as a general purpose IE system which can conform to the MUC task specifications for named entity identification, coreference resolution, IE template element and relation identification, and the construction of scenario-specific IE templates. The system is basically a pipeline architecture consisting of tokenisation, sentence splitting, part-of-speech tagging, morphological stemming, list lookup, parsing with semantic interpretation, proper name matching, and discourse interpretation. The latter stage constructs a discourse model, based on a predefined domain model, using the, often partial, semantic analyses supplied by the parser.</Paragraph>
    <Paragraph position="1"> The domain model represents a hierarchy of domain-relevant concept nodes, together with associated properties. It is expressed in the XI formalism (Gaizauskas, 1995) which provides a basic inheritance mechanism for property values and the ability to represent multiple classificatory dimensions in the hierarchy. Instances of concepts mentioned in a text are added to the domain model, populating it to become a text-, or discourse-, specific model.</Paragraph>
    <Paragraph position="2"> Coreference resolution is carried out by attempting to merge each newly added instance, including pronouns, with instances already present in the model. The basic mechanism is to examine, for each new-old pair of instances: semantic type consistency/similarity in the concept hierarchy; attribute value consistency/similarity, and a set of heuristic rules, some specific to pronouns, which can act to rule out a proposed merge. These rules can refer to various lexical, syntactic, semantic, and positional information about instances. The integration of the focus-based approach replaces the heuristic rules for pronouns, and represents the use of LaSIE as an evaluation platform for more theoretically motivated algorithms. It is possible to extend the approach to include definite NPs but, at present, the existing rules are retained for non-pronominal anaphora in the MUC coreference task: proper names, definite noun phrases and bare nouns.</Paragraph>
  </Section>
  <Section position="6" start_page="74" end_page="76" type="metho">
    <SectionTitle>
4 Implementing Focus-Based
</SectionTitle>
    <Paragraph position="0"> Our implementation makes use of the algorithm proposed in (Azzam, 1996), where elementary events (EEs, effectively simple clauses) are used as basic processing units, rather than sentences.</Paragraph>
    <Paragraph position="1"> Updating the focus registers and the application of interpretation rules (IRs) for pronoun resolution then takes place after each EE, permitting intrasentential references3 In addition, an initial 'expected focus' is determined based on the first EE in a text, providing a potential antecedent for any pronoun within the first EE.</Paragraph>
    <Paragraph position="2"> Development of the algorithm using real-world texts resulted in various further refinements to the algorithm, in both the IRs and the rules for updating the focus registers. The following sections describe the two rules sets separately, though they are highly interrelated in both development and processing.</Paragraph>
    <Section position="1" start_page="74" end_page="75" type="sub_section">
      <SectionTitle>
4.1 Updating the Focus
</SectionTitle>
      <Paragraph position="0"> The algorithm includes two new focus registers, in addition to those mentioned in section 2: AFS, the actor focus stack, used to record previous AF (actor focus) values and so allow a separate set of IRs for agent pronouns (animate verb subjects); and Intra-AFL, the intrasentential alternate focus list, used to record candidate foci from the current EE only.</Paragraph>
      <Paragraph position="1"> In the space available here, the algorithm is best described through an example showing the use of the registers. This example is taken from a New York Times article in the MUC-7 training corpus on aircraft crashes: 1An important limitation of Sidner's algorithm, noted in (Azzam, 1996), is that the focus registers are only updated after each sentence. Thus antecedents proposed for an anaphor in the current sentence will always be from the previous sentence or before and intrasentential references axe impossible.</Paragraph>
      <Paragraph position="2">  State Police said witnesses told them the propeller was not turning as the plane descended quickly toward the highway in Wareham near Exit 2. It hit a tree.</Paragraph>
      <Paragraph position="3"> EE-I: State Police said tell_event An 'expected focus' algorithm applies to initialise the registers as follows:</Paragraph>
      <Paragraph position="5"> Intra-AFL remains empty because EE-1 contains no other candidate foci. No other registers are affected by the expected focus.</Paragraph>
      <Paragraph position="6"> No pronouns occur in EE-1 and so no IRs apply.</Paragraph>
      <Paragraph position="7"> EE-2: witnesses told them The Intra-AFL is first initialised with all (non-pronominal) candidate foci in the EE:</Paragraph>
      <Paragraph position="9"> The IRs are then applied to the first pronoun, them, and, in this case, propose the current AF, State Police, as the antecedent. The Intra-AFL is immediately updated to add the antecedent:</Paragraph>
      <Paragraph position="11"> EE-2 has a pronoun in 'thematic' position, 'theme' being either the object of a transitive verb, or the subject of an intransitive or the copula (following (Gruber, 1976)). Its antecedent therefore becomes the new CF, with the previous value moving to the FS. EE-2 has an 'agent', where this is an animate verb subject (again as in (Gruber, 1976)), and this becomes the new AF. Because the old AF is now the CF, it is not added to the AFS as it would be otherwise. After each EE the Intra-AFL is added to the current AFL, excluding the CF.</Paragraph>
      <Paragraph position="12"> The state after EE-2 is then:</Paragraph>
      <Paragraph position="14"> EE-3: the propeller was not turning The Intra-AFL is reinitialised with candidate foci from this EE:</Paragraph>
      <Paragraph position="16"> No pronouns occur in EE-3 and so no IRs apply. The 'theme', propeller here because of the copula, becomes the new CF and the old one is added to the FS. The AF remains unchanged as the current EE lacks an agent:</Paragraph>
      <Paragraph position="18"> In the current algorithm the AFL is reset at this point, because EE-4 ends the sentence.</Paragraph>
      <Paragraph position="19"> EE-5: it hit a tree Intra-AFL = a tree The IRs resolve the pronoun it with the CF:</Paragraph>
      <Paragraph position="21"/>
    </Section>
    <Section position="2" start_page="75" end_page="76" type="sub_section">
      <SectionTitle>
4.2 Interpretation Rules
</SectionTitle>
      <Paragraph position="0"> Pronouns are divided into three classes, each with a distinct set of IRs proposing antecedents: Personal pronouns acting as agents (animate subjects): (e.g. he in Shotz said he knew the pilots) AF proposed initially, then animate members of AFL.</Paragraph>
      <Paragraph position="1"> Non-agent pronouns: (e.g. them in EE-2 above and it in EE-5) CF proposed initially, then members of the AFL and FS.</Paragraph>
      <Paragraph position="2"> Possessive, reciprocal and reflexive pronouns (PRRs): (e.g. their in the brothers had left and were on their way home) Antecedents proposed from the Intra-AFL, allowing intra-EE references.</Paragraph>
      <Paragraph position="3"> Antecedents proposed by the IRs are accepted or rejected based on their semantic type and feature compatibility, using the semantic and attribute value similarity scores of LaSIE's existing coreference mechanism.</Paragraph>
      <Paragraph position="4"> 5 Evaluation with the MUC Corpora As part of MUC (Grishman and Sundheim, 1996), coreference resolution was evaluated as a sub-task of information extraction, which involved negotiating a definition of coreference relations that could be reliably evaluated. The final definition included only 'identity' relations between text strings: proper nouns, common nouns and pronouns. Other possible coreference relations, such as 'part-whole', and non-text strings (zero anaphora) were excluded.  The definition was used to manually annotate several corpora of newswire texts, using SGML markup to indicate relations between text strings. Automatically annotated texts, produced by systems using the same markup scheme, were then compared with the manually annotated versions, using scoring software made available to MUC participants, based on (Vilain et al., 1995).</Paragraph>
      <Paragraph position="5"> The scoring software calculates the standard Information Retrieval metrics of 'recall' and 'precision', 2 together with an overall f-measure. The following section presents the results obtained using the corpora and scorer provided for MUC-7 training (60 texts, average 581 words per text, 19 words per sentence) and evaluation (20 texts, average 605 words per text, 20 words per sentence), the latter provided for the formal MUC-7 run and kept blind during development.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>