<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1013">
  <Title>Ensemble Methods for Unsupervised WSD</Title>
  <Section position="8" start_page="103" end_page="103" type="concl">
    <SectionTitle>
6 Conclusions and Discussion
</SectionTitle>
    <Paragraph position="0"> In this paper we have presented an evaluation study of four well-known approaches to unsupervised WSD. Our comparison involved type- and token-based disambiguation algorithms relying on different kinds of WordNet relations and different amounts of corpus data. Our experiments revealed two important ndings. First, type-based disambiguation yields results superior to a token-based approach. Using predominant senses is preferable to disambiguating instances individually, even for token-based algorithms. Second, the outputs of the different approaches examined here are suf ciently diverse to motivate combination methods for unsupervised WSD. We de ned several ensembles on the predominant sense outputs of individual methods and showed that combination systems outperformed their best components both on the SemCor and Senseval-3 data sets.</Paragraph>
    <Paragraph position="1"> The work described here could be usefully employed in two tasks: (a) to create preliminary annotations, thus supporting the annotate automatically, correct manually methodology used to provide high volume annotation in the Penn Treebank project; and (b) in combination with supervised WSD methods that take context into account; for instance, such methods could default to an unsupervised system for unseen words or words with uninformative contexts.</Paragraph>
    <Paragraph position="2"> In the future we plan to integrate more components into our ensembles. These include not only domain driven disambiguation algorithms (Strapparava et al., 2004) but also graph theoretic ones (Mihalcea, 2005) as well as algorithms that quantify the degree of association between senses and their co-occurring contexts (Mohammad and Hirst, 2006). Increasing the number of components would allow us to employ more sophisticated combination methods such as unsupervised rank aggregation algorithms (Tan and Jin, 2004).</Paragraph>
  </Section>
class="xml-element"></Paper>