<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1083">
  <Title>Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 660-667, Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Multi-Lingual Coreference Resolution With Syntactic Features</Title>
  <Section position="6" start_page="665" end_page="665" type="evalu">
    <SectionTitle>
5 Related Work
</SectionTitle>
    <Paragraph position="0"> Many researchers have used the syntactic information in their coreference system before. For example, Hobbs (1976) uses a set of rules that are applied to parse trees to determine the antecedent of a pronoun. The rule precedence is determined heuristically and no weight is used.</Paragraph>
    <Paragraph position="1"> Lappin and Leass (1994) extracted rules from the output of the English Slot Grammar (ESG) (McCord, 1993).</Paragraph>
    <Paragraph position="2"> Rule weights are assigned manually and the system resolves the third person pronouns and re exive pronouns only. Ge et al. (1998) uses a non-parametrized statistical model to nd the antecedent from a list of candidates generated by applying the Hobbs algorithm to the English Penn Treebank. Kehler et al. (2004) experiments making use of predicate-argument structure extracted from a large TDT-corpus. Compared with these work, our work uses machine-generated parse trees from which trainable features are extracted in a maximum-entropy coreference system, while (Ge et al., 1998) assumes that correct parse trees are given. Feature weights are automatically trained in our system while (Lappin and Leass, 1994; Stuckardt, 2001) assign weights manually.</Paragraph>
    <Paragraph position="3"> There are a large amount of published work (Morton, 2000; Soon et al., 2001; Ng and Cardie, 2002; Yang et al., 2003; Luo et al., 2004; Kehler et al., 2004) using machine-learning techniques in coreference resolution.</Paragraph>
    <Paragraph position="4"> But none of these work tried to compute complex linguistic concept such as governing category 3 . Our work demonstrates how relevant linguistic knowledge can be derived automatically from system-generated parse trees and encoded into computable and trainable features in a machine-learning framework.</Paragraph>
  </Section>
class="xml-element"></Paper>