<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1025">
  <Title>Exploiting Semantic Role Labeling, WordNet and Wikipedia for Coreference Resolution</Title>
  <Section position="3" start_page="192" end_page="192" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Vieira &amp; Poesio (2000), Harabagiu et al. (2001), and Markert &amp; Nissim (2005) explore the use of WordNet for different coreference resolution subtasks, such as resolving bridging reference, otherand definite NP anaphora, and MUC-style coreference resolution. All of them present systems which infer coreference relations from a set of potential antecedents by means of a WordNet search. Our approach to WordNet here is to cast the search results in terms of semantic similarity measures. Their output can be used as features for a learner. These measures are not specifically developed for coreference resolution but simply taken 'off-the-shelf' and applied to our task without any specific tuning -- i.e.</Paragraph>
    <Paragraph position="1"> in contrast to Harabagiu et al. (2001), who weight WordNet relations differently in order to compute the confidence measure of the path.</Paragraph>
    <Paragraph position="2"> To the best of our knowledge, we do not know of any previous work using Wikipedia or SRL for coreference resolution. In the case of SRL, this layer of semantic context abstracts from the specific lexical expressions used, and therefore represents a higher level of abstraction than (still related) work involving predicate argument statistics. Kehler et al.</Paragraph>
    <Paragraph position="3"> (2004) observe no significant improvement due to predicate argument statistics. The improvement reported by Yang et al. (2005) is rather caused by their twin-candidate model than by the semantic knowledge. Employing SRL is closer in spirit to Ji et al. (2005), who explore the employment of the ACE 2004 relation ontology as a semantic filter.</Paragraph>
  </Section>
class="xml-element"></Paper>