<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1025">
  <Title>Exploiting Semantic Role Labeling, WordNet and Wikipedia for Coreference Resolution</Title>
  <Section position="8" start_page="197" end_page="198" type="concl">
    <SectionTitle>
5 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> The results are somehow surprising, as one would not expect a community-generated categorization to be almost as informative as a well structured  lexical taxonomy such as WordNet. Nevertheless Wikipedia offers promising results, which we expect to improve as well as the encyclopedia goes under further development.</Paragraph>
    <Paragraph position="1"> In this paper we investigated the effects of using different semantic knowledge sources within a machine learning based coreference resolution system. This involved mining the WordNet taxonomy and the Wikipedia encyclopedic knowledge base, as well as including semantic parsing information, in order to induce semantic features for coreference learning. Empirical results show that coreference resolution benefits from semantics. The generated model is able to learn selectional preferences in cases where surface morpho-syntactic features do not suffice, i.e. pronoun and common name resolution. While the results given by using 'the free encyclopedia that anyone can edit' are satisfactory, major improvements can come from developing efficient query strategies - i.e. a more refined disambiguation technique taking advantage of the context in which the queries (e.g. referring expressions) occur.</Paragraph>
    <Paragraph position="2"> Future work will include turning Wikipedia into an ontology with well defined taxonomic relations, as well as exploring its usefulness of for other NLP applications. We believe that an interesting aspect of Wikipedia is that it offers large coverage resources for many languages, thus making it a natural choice for multilingual NLP systems.</Paragraph>
    <Paragraph position="3"> Semantics plays indeed a role in coreference resolution. But semantic features are expensive to  compute and the development of efficient methods is required to embed them into large scale systems. Nevertheless, we believe that exploiting semantic knowledge in the manner we described will assist the research on coreference resolution to overcome the plateauing in performance observed by Kehler et al. (2004).</Paragraph>
    <Paragraph position="4"> Acknowledgements: This work has been funded by the Klaus Tschira Foundation, Heidelberg, Germany. The first author has been supported by a KTF grant (09.003.2004). We thank Katja Filippova, Margot Mieskes and the three anonymous reviewers for their useful comments.</Paragraph>
  </Section>
class="xml-element"></Paper>