<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2503">
  <Title>Relating WordNet Senses for Word Sense Disambiguation</Title>
  <Section position="8" start_page="44" end_page="44" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have investigated methods for relating Word-Net word senses based on distributionally similar nearest neighbours and using the JCN measure.</Paragraph>
    <Paragraph position="1"> Whilst the senses foragiven word can beclustered into sense groups, we propose the use of ranked lists to relate the senses of a word to each other.</Paragraph>
    <Paragraph position="2"> In this way, the granularity can be determined for a given application and the appropriate number of senses for a given word is not needed a priori. We have encouraging results for nouns when comparing RLISTs to manually created gold-standards.</Paragraph>
    <Paragraph position="3"> Wehave produced a newgold-standard forevaluation based on the words used in SEVAL-2 ENG LEX. Wedid this because there is noavailable documentation on inter-annotator agreement for the SEGR. In future, we hope to produce another gold-standard resource where the informants indicate a degree of relatedness, rather than a binary choice of related or unrelated for each pair.</Paragraph>
    <Paragraph position="4"> We would like to see the impact that coarser-grained WSD has on a task or application. Given the lack of a plug and play application for feeding disambiguated data, we hope to examine the benefits on some lexical acquisition tasks that might feed into an application, for example sense ranking (McCarthy et al., 2004) or selectional preference acquisition.</Paragraph>
    <Paragraph position="5"> At this stage we have only experimented with nouns, we hope to go on relating senses in other parts-of-speech, particularly verbs since they have very fine-grained distinctions in WordNet and many of the subtler distinctions are quite probably not important for some applications. (Palmer et al., forthcoming) has clearly demonstrated the necessity for using predicate-argument structure when grouping verb senses, so we want to exploit such information for verbs.</Paragraph>
    <Paragraph position="6"> We have focused on improving the first sense heuristic, but we plan to use our groupings with context-based WSD. To avoid a requirement for  hand-tagged training data, we plan to exploit the collocates of nearest neighbours.</Paragraph>
  </Section>
class="xml-element"></Paper>