<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1022">
  <Title>Coarse-to-fine n-best parsing and MaxEnt discriminative reranking</Title>
  <Section position="7" start_page="178" end_page="179" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> This paper has described a dynamic programming n-best parsing algorithm that utilizes a heuristic coarse-to-fine refinement of parses. Because the coarse-to-fine approach prunes the set of possible parse edges beforehand, a simple approach which enumerates the n-best analyses of each parse edge is not only practical but quite efficient.</Paragraph>
    <Paragraph position="1"> We use the 50-best parses produced by this algorithm as input to a MaxEnt discriminative reranker. The reranker selects the best parse from this set of parses using a wide variety of features. The system we described here has an f-score of 0.91 when trained and tested using the standard PARSEVAL framework.</Paragraph>
    <Paragraph position="2"> This result is only slightly higher than the highest reported result for this test-set, Bod's (.907) (Bod, 2003). More to the point, however, is that the system we describe is reasonably efficient so it can be used for the kind of routine parsing currently being handled by the Charniak or Collins parsers.</Paragraph>
    <Paragraph position="3"> A 91.0 f-score represents a 13% reduction in f- null measure error over the best of these parsers.2 Both the 50-best parser, and the reranking parser can be found at ftp://ftp.cs.brown.edu/pub/nlparser/, named parser and reranker respectively.</Paragraph>
    <Paragraph position="4"> Acknowledgements We would like to thanks Michael Collins for the use of his data and many helpful comments, and Liang Huang for providing an early draft of his paper and very useful comments on our paper. Finally thanks to the National Science Foundation for its support (NSF IIS-0112432, NSF 9721276, and NSF DMS-0074276).</Paragraph>
  </Section>
class="xml-element"></Paper>