<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1023">
  <Title>Coreference Resolution Using Competition Learning Approach</Title>
  <Section position="3" start_page="0" end_page="3" type="intro">
    <SectionTitle>
2 The Single-Candidate Model
</SectionTitle>
    <Paragraph position="0"> The main idea of the single-candidate model for coreference resolution is to recast the resolution as a binary classification problem.</Paragraph>
    <Paragraph position="1"> During training, a set of training instances is generated for each anaphor in an annotated text.</Paragraph>
    <Paragraph position="2"> An instance is formed by the anaphor and one of its antecedent candidates. It is labeled as positive or negative based on whether or not the candidate is tagged in the same coreferential chain of the anaphor.</Paragraph>
    <Paragraph position="3"> After training, a classifier is ready to resolve the NPs  encountered in a new document. For each NP under consideration, every one of its antecedent candidates is paired with it to form a test instance. The classifier returns a number between 0 and 1 that indicates the likelihood that the candidate is coreferential to the NP.</Paragraph>
    <Paragraph position="4"> The returned confidence value is commonly used as the competition criterion to rank the candidate. Normally, the candidates with confidences less than a selection threshold (e.g. 0.5) are discarded. Then some algorithms are applied to choose one of the remaining candidates, if any, as the antecedent. For example, &amp;quot;Closest-First&amp;quot; (Soon et al., 2001) selects the candidate closest to the anaphor, while &amp;quot;Best-First&amp;quot; (Aone and Bennett, 1995; Ng and Cardie, 2002a) selects the candidate with the maximal confidence value.</Paragraph>
    <Paragraph position="5"> One limitation of this model, however, is that it only considers the relationships between a NP encountered and one of its candidates at a time during its training and testing procedures. The confidence value reflects the probability that the candidate is coreferential to the NP in the overall  In this paper a NP corresponds to a Markable in MUC coreference resolution tasks.</Paragraph>
    <Paragraph position="6"> distribution  , but not the conditional probability when the candidate is concurrent with other competitors. Consequently, the confidence values are unreliable to represent the true competition criterion for the candidates.</Paragraph>
    <Paragraph position="7"> To illustrate this problem, just suppose a data set where an instance could be described with four exclusive features: F1, F2, F3 and F4. The ranking of candidates obeys the following rule:</Paragraph>
    <Paragraph position="9"> didates with the feature Fi on. The mark of &amp;quot;&gt;&gt;&amp;quot; denotes the preference relationship, that is, the candidates in CS  is preferred to those in CS</Paragraph>
    <Paragraph position="11"> denote the class value of a leaf node &amp;quot;F2 = 1&amp;quot; and &amp;quot;F3 = 1&amp;quot;, respectively. It is possible that CF</Paragraph>
  </Section>
class="xml-element"></Paper>