<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1119">
  <Title>A Statistical Approach to Anaphora Resolution</Title>
  <Section position="10" start_page="169" end_page="169" type="concl">
    <SectionTitle>
8 Conclustion and Future Research
</SectionTitle>
    <Paragraph position="0"> We have presented a statistical method for pronominal anaphora that achieves an accuracy of 84.2%. The main advantage of the method is its essential simplicity. Except for implementing the Hobbs referent-ordering algorithm, all other system knowledge is imbedded in tables giving the various component probabilities used in the probability model.</Paragraph>
    <Paragraph position="1"> We believe that this simplicity of method will translate into comparative simplicity as we improve the method. Since the research described herein we have thought of other influences on anaphora resolution and their statistical correlates. We hope to include some of them in future work.</Paragraph>
    <Paragraph position="2"> Also, as indicated by the work on unsupervised learning of gender information, there is a growing arsenal of learning techniques to be applied to statistical problems. Consider again the three high-salience words to which our unsupervised learning program assigned incorrect gender: &amp;quot;husband&amp;quot;, &amp;quot;wife&amp;quot;, and &amp;quot;years.&amp;quot; We suspect that had our pronoun-assignment method been able to use the topic information used in the complete method, these might well have been decided correctly. That is, we suspect that &amp;quot;husband&amp;quot;, for example, was decided incorrectly because the topic of the article was the woman, there was a mention of her &amp;quot;husband,&amp;quot; but the article kept on talking about the woman and used the pronoun &amp;quot;she.&amp;quot; While our simple program got confused, a program using better statistics might not have. This too is a topic for future research.</Paragraph>
  </Section>
class="xml-element"></Paper>