<?xml version="1.0" standalone="yes"?>
<Paper uid="P89-1015">
  <Title>ACQUIRING DISAMBIGUATION RULES FROM TEXT</Title>
  <Section position="8" start_page="124" end_page="124" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> I have described a training algorithm that uses an existing deterministic parser together with a corpus of tagged text to acquiring rules for disambiguating lexical category. Performance of the trained set of rules is much better than the previous hand-written rule set (error rate reduced by half). The success of the disambiguation procedure depends on the linguistic knowledge embodied in the parser in a number of ways.</Paragraph>
    <Paragraph position="1"> It uses the data structures and linguistic categories of the parser, focusing the rule acquisition mechanism on relevant elements.</Paragraph>
    <Paragraph position="2"> It is embedded in the parsing process so that parser actions can set things up for acquisition (for example, adverbs axe in effect removed within elements of the auxiliary, restoring the contiguity of auxiliary elements). null It uses the grammar rules to identify words that are grammatically related, and are therefore relevant to disambiguation.</Paragraph>
    <Paragraph position="3"> It can use rough models of complementation and modification to help identify words that are related.</Paragraph>
    <Paragraph position="4"> Finally, the parser always provides a default action. This permits the incremental improvement of the parser, since it can take advantage of more specific information when it is available, but it will always disambiguate somehow, no matter whether it has acquired the appropriate rules or not.</Paragraph>
    <Paragraph position="5"> This work demonstrates the feasibility of acquiring the linguistic information needed to analyze unrestricted text from text itself. Further improvements in syntactic analyzers will depend on such automatic acquisition of grammatical and lexical facts.</Paragraph>
  </Section>
class="xml-element"></Paper>