<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0802">
  <Title>Sense clusters for Information Retrieval: Evidence from Semcor and the EuroWordNet InterLingual Index</Title>
  <Section position="6" start_page="17" end_page="17" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> We examined three different types of sense clustering criteria with an Information Retrieval application in mind: methods based on the word-net structure (such as generalization, cousins, sisters...); co-occurrence of senses obtained from Semcor; and equivalent translations of senses in other languages via the EuroWordNet InterLingual Index (ILI). We conclude that a) different NLP applications demand not only different sense granularities but different (possibly overlapped) sense elusterings, b) co-occurrence of senses in Semcor provide strong evidence for Information Retrieval clusters, unlike methods based on word-net structure and systematic polysemy, c) parallel polysemy in two or more languages via the ILI, besides providing sense clusters for MT and CLIR, is correlated with coocurring senses in Semcor, and thus can be useful to obtain IR dusters as well.</Paragraph>
    <Paragraph position="1"> Both approaches to IR clusters fbr WN/EWN (evidence from Semcor and from the ILl) seem very promising. The positive and negative evidence from SeIncor (above 500 clusters each) can possibly be used in a Machine Learning approach to find additional dusters for the rem~inlng sense distinctions without enough evidence from Serecot. The parallel polysemy criteria, over EWN, is highly productive (more than one candidate per word in our experiments), although a more diverse set of languages would probably produce a higher rate of valid clusters.</Paragraph>
  </Section>
class="xml-element"></Paper>