<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-1029">
  <Title>A Tool for Automated Revision of Grammars for NLP Systems</Title>
  <Section position="6" start_page="214" end_page="216" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented a set of algorithms and an interactive tool for automatically revising grammars of NLP systems to disallow identified counter-examples (sentences or sets of sentences accepted by the current grammar but deemed to be irrelevant for a given application). We have successfully used the tool to constrain overgeneralizing grammars of speech understanding systems and obtained 20-30% higher recognition accuracy. However, we believe the primary benefit of using our tool is the tremendously reduced effort for the grammar developer. Our technique relieves the grammar developer from the burden of going through the tedious and time consuming task of revising grammars by manually modifying production rules one at a time. Instead, the grammar developer simply identifies counter-examples to an interactive tool that revises the grammar to invalidate the identified sentences.</Paragraph>
    <Paragraph position="1"> We also discussed an MDL based algorithm for grammar compaction to reduce the size of the revised grammar. Thus, using a combination of the algorithms presented in this paper, one can obtain a compact grammar that is guaranteed to disallow the counter-examples.</Paragraph>
    <Paragraph position="2"> Although our discussion here was focussed on speech understanding applications, the algorithms and the tool described here are applicable for any domain where grammars are used. We are currently implementing an extension of the grammar modifier to handle attribute-value grammars. We outlined an  approach for automated modification of attribute-value grammars in Section 3.</Paragraph>
    <Paragraph position="3"> We conclude that algorithms for automatically constraining grammars based on counter-examples can be highly effective in reducing the burden on grammar developers to develop constrained, domain specific grammars.</Paragraph>
    <Paragraph position="4"> Moreover, these algorithms can be used in any applications, which deal with grammars.</Paragraph>
  </Section>
class="xml-element"></Paper>