<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0833">
  <Title>Simple Features for Statistical Word Sense Disambiguation</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> There is no xed context window size applicable to all ambiguous words in the Na ve Bayes approach: keeping a large context window provides domain information which increases the resolution accuracy for some target words but not others. For non-topical words, large window size is selected only in order to exploit the distribution of samples.</Paragraph>
    <Paragraph position="1"> line is really di cult, 2) When the di erence is mostly on the commonest sense being seen more than expected, so the score is favored (7 words out of 57 satisfy these conditions.) Rough syntactic information performed well in our second system using Maximum Entropy modeling. This suggests that some senses can be strongly identi ed by syntax, leaving resolution of other senses to other methods. A simple, rough heuristic for recognizing when to rely on syntactic information in our system is when the selected window size by Na ve Bayes is relatively small.</Paragraph>
    <Paragraph position="2"> We tried two simple methods for combining the two methods: considering context words as features in Max Entropy learner, and, establishing a separate Na ve Bayes learner for each syntactic/semantic feature and adding their scores to the basic contextual Na ve Bayes. These preliminary experiments did not result in any noticeable improvement.</Paragraph>
    <Paragraph position="3"> Finally, using more semantic features from WordNet, such as verb sub-categorization frames (which are not consistently available) may help in distinguishing the senses.</Paragraph>
  </Section>
class="xml-element"></Paper>