<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3814">
  <Title>Evaluating and optimizing the parameters of an unsupervised graph-based WSD algorithm</Title>
  <Section position="9" start_page="95" end_page="95" type="concl">
    <SectionTitle>
7 Conclusions and further work
</SectionTitle>
    <Paragraph position="0"> This paper has explored two sides of HyperLex: the optimization of the free parameters, and the empirical comparison on a standard benchmark against other WSD systems. We use hand-tagged corpora to map the induced senses to WordNet senses.</Paragraph>
    <Paragraph position="1"> Regarding the optimization of parameters, we used a another testbed (S2LS) comprising different words to select the best parameter. We consistently improve the results of the parameters by V'eronis, which is not perhaps so surprising, but the method allows to fine-tune the parameters automatically to a given corpus given a small test set.</Paragraph>
    <Paragraph position="2"> Comparing unsupervised systems against supervised systems is seldom done. Our results indicate that HyperLex with the supervised mapping is on par with a state-of-the-art system which uses bag-of-words features only. Given the information loss inherent to any mapping, this is an impressive result. The comparison to other unsupervised systems is difficult, as each one uses a different mapping strategy and a different amount of supervision.</Paragraph>
    <Paragraph position="3"> For the future, we would like to look more closely the micro-senses induced by HyperLex, and see if we can group them into coarser clusters. We also plan to apply the parameters to the Senseval 3 all-words task, which seems well fit for HyperLex: the best supervised system only outperforms MFS by a few points in this setting, and the training corpora used (Semcor) is not related to the test corpora (mainly Wall Street Journal texts).</Paragraph>
    <Paragraph position="4"> Graph models have been very successful in some settings (e.g. the PageRank algorithm of Google), and have been rediscovered recently for natural language tasks like knowledge-based WSD, textual entailment, summarization and dependency parsing.</Paragraph>
    <Paragraph position="5"> We would like to test other such algorithms in the same conditions, and explore their potential to integrate different kinds of information, especially the local or syntactic features so successfully used by supervised systems, but also more heterogeneous information from knowledge bases.</Paragraph>
  </Section>
class="xml-element"></Paper>