<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1023">
  <Title>Coreference Resolution Using Competition Learning Approach</Title>
  <Section position="12" start_page="10" end_page="10" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> In this paper we have proposed a competition learning approach to coreference resolution. We started with the introduction of the single-candidate model adopted by most supervised machine learning approaches. We argued that the confidence values returned by the single-candidate classifier are not reliable to be used as ranking criterion for antecedent candidates. Alternatively, we presented a twin-candidate model that learns the competition criterion for antecedent candidates directly. We introduced how to adopt the twin-candidate model in our competition learning approach to resolve the coreference problem. Particularly, we proposed a candidate filtering algorithm that can effectively reduce the computational cost and data noises.</Paragraph>
    <Paragraph position="1"> The experimental results have proved the effectiveness of our approach. Compared with the base-line approach using the single-candidate model, the F-measure increases by 1.9 and 1.5 for MUC-6 and MUC-7 data set, respectively. The gains in the pronoun resolution contribute most to the overall improvement of coreference resolution.</Paragraph>
    <Paragraph position="2"> Currently, we employ the single-candidate classifier to filter the candidate set during resolution. While the filter guarantees the qualification of the candidates, it removes too many positive candidates, and thus the recall suffers. In our future work, we intend to adopt a looser filter together with an anaphoricity determination module (Bean and Riloff, 1999; Ng and Cardie, 2002b). Only if an encountered NP is determined as an anaphor, we will select an antecedent from the candidate set generated by the looser filter. Furthermore, we would like to incorporate more syntactic features into our feature set, such as grammatical role or syntactic parallelism. These features may be helpful to improve the performance of pronoun resolution. null</Paragraph>
  </Section>
class="xml-element"></Paper>