<?xml version="1.0" standalone="yes"?>
<Paper uid="W99-0634">
  <Title>Corpus-Based Learning for Noun Phrase Coreference Resolution</Title>
  <Section position="8" start_page="289" end_page="289" type="relat">
    <SectionTitle>
4 Related Work
</SectionTitle>
    <Paragraph position="0"> There is a long tradition of work on coreference resolution within computational linguistics, but most of them are not subjected to empirical evaluation until recently. Among the work that reported quantitative evaluation results, most are not based on learning from an annotated corpus (Baldwin, 1997; Kameyama, 1997; Lappin and Leass, 1994; Mitkov, 1997).</Paragraph>
    <Paragraph position="1"> To our knowledge, the work of (Aone and Bennett, 1995; Ge et al., 1998; McCarthy and Lehnert, 1995) are the only ones that are based on learning from an annotated corpus. Ge et al. (Ge et al., 1998) used a statistical model for resolving pronouns. In contrast, we used a decision tree learning algorithm and resolved general noun phrases, not just pronouns. Both the work of (Aone and Bennett, 1995; Mc-Carthy and Lehnert, 1995) employed decision tree learning. However, the features they used include domain-specific ones like DNP-F (definite NP whose referent is a facility) (Aone and Bennett, 1995), JV-CHILD-i (does i refer to a joint venture formed as the result of a tie-up) (McCarthy and Lehnert, 1995), etc.</Paragraph>
    <Paragraph position="2"> In contrast, all our 10 features are domainindependent, which makes our coreference engine a domain-independent module. Moreover, both (Aone and Bennett, 1995) and (McCarthy and Lehnert, 1995) made simplifying assumptions in their experimental evaluations. Since the accuracy of coreference resolution relies on the correct identification of the candidate noun phrases, both (Aone and Bennett, 1995) and (McCarthy and Lehnert, 1995) only evaluated their systems on noun phrases that have been correctly identified. In contrast, we evaluated our coreference resolution engine as part of a total system which has to first identify all the candidate noun phrases and has to deal with the inevitable noisy data when mistakes occur in noun phrase identification. Also, the evaluation of (Aone and Bennett, 1995) and (McCarthy and Lehnert, 1995) only focused on specific types of noun phrases (organizations and business entities), and (Aone and Bennett, 1995) dealt only with Japanese texts. Our evaluation was done on all types of English noun phrases instead.</Paragraph>
    <Paragraph position="3"> None of the systems in MUC-7 adopted a learning approach to coreference resolution (Chinchor, 1998). Among the MUC-6 systems, the only one that we can directly compare to is the UMass system, which also used C4.5 for coreference resolution. The other MUC-6 systems were not based on a learning approach.</Paragraph>
    <Paragraph position="4"> The score of the UMass system is not high compared to the rest of the MUC-6 systems. In particular, the system's recall is relatively low. As explained in (Fisher et al., 1995), the reason for this is that it only concentrated on coreference relationships among references to people and organizations. Our system, as opposed to the UMass system, considered all types of markables. The score of our system is higher than that of the UMass system, and the difference is statistically significant at p = 0.05. Thus, the contribution of our work is in showing that a learning approach, when evaluated on a common coreference data set, is able to achieve accuracy competitive with state-of-the-art systems using non-learning approaches.</Paragraph>
  </Section>
class="xml-element"></Paper>