<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-1028">
  <Title>Experiments with Corpus-based LFG Specialization</Title>
  <Section position="7" start_page="207" end_page="207" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> Sophisticated grammar formalisms are very useful and convenient when designing high-coverage grammars for natural languages. Very expressive grammatical constructs can make the task of developing and maintaining such a large resource considerably easier. On the other hand, their use can result in a considerable increase in grammatical ambiguity. Gramnaar-compilation techniques based on grammar structure alone are insufficient remedies in those cases, as they cannot access the information required to determine which alternatives to retain and which alternatives to discard.</Paragraph>
    <Paragraph position="1"> The current article demonstrates that a relatively simple pruning technique, employing the kind of reference corpus that is typically used for grammar development and thus often already available, can significantly improve parsing performance. On large lexical functional grammars, speedups of up to a factor 6 were observed, at the price of a. reduction in grammatical coverage of about 13%. A simple two-stage architecture was also proposed that preserves the anyparse measure of the original grammar, demonstrating that significant speedups can be obtained without increasing the number of parsing failures.</Paragraph>
    <Paragraph position="2"> Future work includes extending the study of corpus-based grammar specialization from first-order grammar pruning to higher-order grammar pruning, thus extending previous work on explanation-based learning for parsing, aad applying it to the LFG fornaalism.</Paragraph>
  </Section>
class="xml-element"></Paper>