<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1037">
  <Title>Guiding a Constraint Dependency Parser with Supertags</Title>
  <Section position="8" start_page="294" end_page="295" type="concl">
    <SectionTitle>
6 Conclusions and future work
</SectionTitle>
    <Paragraph position="0"> We have shown that a statistical supertagging component can significantly improve the parsing accuracy of a general-purpose dependency parser for German. The error rate among syntactic attachments can be reduced by 24% over an already competitive baseline. After all, the integration of the supertagging results helped to reach a quality level which compares favourably with the state-of-the-art in probabilistic dependency parsing for German as defined with 87.34%/90.38% labelled/unlabelled attachment accuracy on this years shared CoNLL task by (McDonald et al., 2005) (see (Foth and Menzel, 2006) for a more detailed comparison). Although the statistical model used in our system is rather simple-minded, it clearly captures at least some distributional char- null acteristics of German text that the hand-written rules do not.</Paragraph>
    <Paragraph position="1"> A crucial factor for success is the defeasible integration of the supertagging predictions via soft constraints. Rather than pursuing a strict filtering approach where supertagging errors are partially compensated by an n-best selection, we commit to only one supertag per word, but reduce its influence. Treating supertag predictions as weak preferences yields the best results. By measuring the accuracy of the different types of predictions made by complex supertags, different weights could also be assigned to the six new constraints.</Paragraph>
    <Paragraph position="2"> Of the investigated supertag models, the most complex ones guide the parser best, although their own accuracy is not the best one, even when measured by the more pertinent component accuracy. Since purely statistical parsing methods do not reach comparable parsing accuracy on the same data, we assume that this trend does not continue indefinitely, but would stop at some point, perhaps already reached.</Paragraph>
  </Section>
class="xml-element"></Paper>