<?xml version="1.0" standalone="yes"?>
<Paper uid="P04-2010">
  <Title>A Machine Learning Approach to German Pronoun Resolution</Title>
  <Section position="3" start_page="0" end_page="0" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Although extensive research has been conducted on statistical anaphora resolution, the bulk of the work has concentrated on the English language. Nevertheless, comparing different strategies helped shape the system described in this paper. null (McCarthy and Lehnert, 1995) were among the first to use machine learning for coreference resolution. RESOLVE was trained on data from MUC-5 English Joint Venture (EJV) corpus and used the C4.5 decision tree algorithm (Quinlan, 1993) with eight features, most of which were tailored to the joint venturte domain. The system achieved an F-measure of 86.5 for full coreference resolution (no values were given for pronouns).</Paragraph>
    <Paragraph position="1"> Although a number this high must be attributed to the specific textual domain, RESOLVE also out-performed the authors' rule-based algorithm by 7.6 percentage points, which encouraged further reseach in this direction.</Paragraph>
    <Paragraph position="2"> Unlike the other systems presented in this section, (Morton, 2000) does not use a decision tree algorithm but opts instead for a maximum entropy model. The model is trained on a subset of the Wall Street Journal, comprising 21 million tokens. The reported F-measure for pronoun resolution is 81.5. However, (Morton, 2000) only attempts to resolve singular pronouns, and there is no mention of what percentage of total pronouns are covered by this restriction.</Paragraph>
    <Paragraph position="3"> (Soon et al., 2001) use the C4.5 algorithm with a set of 12 domain-independent features, ten syntactic and two semantic. Their system was trained on both the MUC-6 and the MUC-7 datasets, for which it achieved F-scores of 62.6 and 60.4, respectively. Although these results are far worse than the ones reported in (McCarthy and Lehnert, 1995), they are comparable to the best-performing rule-based systems in the respective competitions. As (McCarthy and Lehnert, 1995), (Soon et al., 2001) do not report separate results for pronouns. (Ng and Cardie, 2002) expanded on the work of (Soon et al., 2001) by adding 41 lexical, semantic and grammatical features. However, since using this many features proved to be detrimental to performance, all features that induced low precision rules were discarded, leaving only 19.</Paragraph>
    <Paragraph position="4"> The final system outperformed that of (Soon et al., 2001), with F-scores of 69.1 and 63.4 for MUC-6 and MUC-7, respectively. For pronouns, the reported results are 74.6 and 57.8, respectively.</Paragraph>
    <Paragraph position="5"> The experiment presented in (Strube et al., 2002) is one of the few dealing with the application of machine learning to German coreference resolution covering definite noun phrases, proper names and personal, possessive and demonstrative pronouns. The research is based on the Heidelberg Text Corpus (see Section 4), which makes it ideal for comparison with our system. (Strube et al., 2002) used 15 features modeled after those used by state-of-the-art resolution systems for English.</Paragraph>
    <Paragraph position="6"> The results for personal and possessive pronouns are 82.79 and 84.94, respectively.</Paragraph>
  </Section>
class="xml-element"></Paper>