<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1023">
  <Title>Coreference Resolution Using Competition Learning Approach</Title>
  <Section position="11" start_page="10" end_page="10" type="relat">
    <SectionTitle>
6 Related Work
</SectionTitle>
    <Paragraph position="0"> A similar twin-candidate model was adopted in the anaphoric resolution system by Connolly et al.</Paragraph>
    <Paragraph position="1"> (1997). The differences between our approach and theirs are: (1) In Connolly et al.'s approach, all the preceding NPs of an anaphor are taken as the antecedent candidates, whereas in our approach we use candidate filters to eliminate invalid or irrelevant candidates.</Paragraph>
    <Paragraph position="2"> (2) The antecedent identification in Connolly et al.'s approach is to apply the classifier to successive pairs of candidates, each time retaining the better candidate. However, due to the lack of strong assumption of transitivity, the selection procedure is in fact a greedy search. By contrast, our approach evaluates a candidate according to the times it wins over the other competitors. Comparatively this algorithm could lead to a better solution.</Paragraph>
    <Paragraph position="3"> (3) Our approach makes use of more indicative features, such as Appositive, Name Alias, String-matching, etc. These features are effective especially for non-pronoun resolution.</Paragraph>
  </Section>
class="xml-element"></Paper>