<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3224">
  <Title>A Distributional Analysis of a Lexicalized Statistical Parsing Model</Title>
  <Section position="9" start_page="0" end_page="0" type="concl">
    <SectionTitle>
8 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> With so many parameters, a lexicalized statistical parsing model seems like an intractable behemoth.</Paragraph>
    <Paragraph position="1"> However, as statisticians have long known, an excellent angle of attack for a mass of unruly data is exploratory data analysis. This paper presents some of the first data visualizations of parameters in a parsing model, and follows up with a numerical analysis of properties of those distributions. In the course of this analysis, we have focused in on the question of bilexical dependencies. By constrainparsing the parser's own output, and by hypothesizing and testing for distributional similarity, we have presented evidence that finally explains that (a) bilexical statistics are actually getting used with great frequency in the parse theories that will ultimately have the highest score, but (b) the distributions involving bilexical statistics are so similar to their back-o counterparts as to make them nearly indistinguishable insofar as making di erent parse decisions. Finally, our analysis has provided for the first time an e ective way to do parameter selection with a generative lexicalized statistical parsing model.</Paragraph>
    <Paragraph position="2"> Of course, there is still much more analysis, hypothesizing, testing and extrapolation to be done. A thorough study of the highest-entropy distributions should reveal new ways in which to use grammar transforms or develop features to reduce the entropy and increase parse accuracy. A closer look at the low-entropy distributions may reveal additional reductions in the size of the model, and, perhaps, a way to incorporate hard constraints without disturbing the more ambiguous parts of the model more suited to machine learning than human engineering.</Paragraph>
  </Section>
class="xml-element"></Paper>