<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1047">
  <Title>A New Approach to Word Sense Disambiguation</Title>
  <Section position="7" start_page="246" end_page="246" type="concl">
    <SectionTitle>
6. CONCLUSIONS AND FUTURE
WORK
</SectionTitle>
    <Paragraph position="0"> In'this paper, we have presented and evaluated models created according to a schema that provides a description of the joint distribution of the values of sense tags and contextual features that is potentially applicable to a wide range of content words. The models were evaluated through a series of experiments that provided the following information: 1) performance results (precision, recall, and total percent correct) for probabilistic classifiers using models created in azcordance with the schema and applied to the disambiguation of several difficult test words; 2) identification of upper and lower bounds for the performance of any probabilistic word-sense classifier using the contextual features defined in the model schema; and 3) a comparison of the performance of classifiers using models generated per the schema to that of classifiers using models selected as described in section 2. The results of these experiments suggest that the model schema is particularly well suited to nouns but that it is also applicable to words in other syntactic categories.</Paragraph>
    <Paragraph position="1"> We feel that the results presented in this paper are encouraging and plan to continue testing the model schema on other words. But it is unreasonable to continue generating over 1,000 manually sense-tagged examples of each word to be disambiguated, as is required if parameters are estimated as we did here. In answer to this problem, other means of parameter estimation are being investigated, including a procedure for obtaining maximum likelihood estimates from untagged data. The procedure is a variant of the EM algorithm (\[7\]) specifically applicable to models of the form described in this paper.</Paragraph>
    <Paragraph position="2"> ACKNOWLEDGEMENTS. The authors would like to gratefully acknowledge the contributions of the following people to the work presented in this paper: Rufus and Beverly Bruce for their help in sense-tagging data, Gerald Rogers for sharing his expertise in statistics, and Ted Dunning for advice and support in all matters having to do with software development.</Paragraph>
  </Section>
class="xml-element"></Paper>