<?xml version="1.0" standalone="yes"?>
<Paper uid="P04-1080">
  <Title>Learning Word Senses With Feature Selection and Order Identification Capabilities</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> Our word sense learning algorithm combined two novel ingredients: feature selection and order identification. Feature selection was formalized as a constrained optimization problem, the output of which was a set of important features to determine word senses. Both cluster structure and cluster number were estimated by minimizing a MDL criterion. Experimental results showed that our algorithm can retrieve important features, estimate cluster number automatically, and achieve better performance in terms of average accuracy than CGD algorithm which required cluster number as input.</Paragraph>
    <Paragraph position="1"> Our word sense learning algorithm is unsupervised in two folds: no requirement of sense tagged data, and no requirement of predefinition of sense number, which enables the automatic discovery of word senses from free text.</Paragraph>
    <Paragraph position="2"> In our algorithm, we treat bag of words in local contexts as features. It has been shown that local collocations and morphology of target word play important roles in word sense disambiguation or discrimination (Leacock et al., 1998; Widdows, 2003). It is necessary to incorporate these more structural information to improve the performance of word sense learning.</Paragraph>
  </Section>
class="xml-element"></Paper>