<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0813">
  <Title>Combining Contextual Features for Word Sense Disambiguation</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have demonstrated that our approach to disambiguating verb senses using maximum entropy models to combine as many linguistic knowledge sources as possible, yields state-of-the-art performance for English. This may be a language-dependent feature, as other experiments indicate that additional linguistic pre-processing does not necessarily improve tagging accuracy for languages like Chinese (Dang et al., 2002).</Paragraph>
    <Paragraph position="1"> In examining the instances that proved troublesome to both the human taggers and the automatic system, we found errors that were tied to subtle sense distinctions which were reconciled by backing off to the more coarse-grained sense groups.</Paragraph>
    <Paragraph position="2"> Achieving higher inter-annotator agreement is necessary in order to provide consistent training data for supervised WSD systems. Lexicographers have long recognized that many natural occurrences of polysemous words are embedded in underspecified contexts and could correspond to more than one specific sense. Annotators need the option of selecting, as an alternative to an explicit sense, either a group of specific senses or a single, broader sense, where specific meaning nuances are subsumed. Sense grouping, already present in a limited way in Word-Net's verb component, can be guided and enhanced by the analysis of inter-annotator disagreements and the development of explicit sense distinction criteria that such an analysis provides.</Paragraph>
  </Section>
class="xml-element"></Paper>