<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0853">
  <Title>A Gloss-centered Algorithm for Disambiguation</Title>
  <Section position="11" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 English Lexical Sample Task
</SectionTitle>
    <Paragraph position="0"> The results of our gloss based disambiguation system show that an optimal configuration of the parameters is essential to get good results. Hyper-Desc(a9 ) glosses together with stemming seem to almost always give better results than other. But it may be worthwhile finding out the weight-age for different types of glosses and use all of them together. However - the algorithm performs better than the baseline algorithm, it still falls short of a decent precision that is generally a pre-requisite for the use of WSD in Machine Translation - a0 a9 %. One obvious reason for this is that no matter how we try  to use WordNet, the descriptive glosses of Word-Net are very sparse and contain very few contextual clues for sense disambiguation. In the task of English Lexical Sample, we further develop the algorithm describe for the previous task and use relatively dense glosses from the training set. The large size of the glosses require us to modify the architecture for ranking glosses. We use and inverted index for indexing the glosses and treat the context of the word to be disambiguated as a query. The senses of the word are ranked using the same set of parameters as described for the earlier task.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.1 Experiments
</SectionTitle>
      <Paragraph position="0"> For this task, the gloss for a word-sense is generated by concatenating the contexts of all training instances for that word-sense. An inverted index is generated for the glosses. The context for a test instance is fired as a query and the senses for the word are ranked using the tf-igf based cosine similarity metric described in section 3.1. The top sense is picked.</Paragraph>
      <Paragraph position="1"> The baseline precision obtained for this task was 53.5% The precision obtained using fine-grained scoring was 66.1% and the recall was 65.7%. The precision obtained using coarse-grained scoring was 74.3% and the recall was 73.9%.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.2 Conclusion
</SectionTitle>
      <Paragraph position="0"> We see that densely populated glosses do help in getting a better precision score. One possible course of action that this finding suggests is some kind of interactive WSD where the user is allowed to correct machine generated tags for some dataset. The contexts for words in the correctly tagged data could next get appended to existing gloss of the corresponding word-sense.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>