<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1015">
  <Title>Combining Deep and Shallow Approaches in Parsing German</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> The paper has presented two approaches to German parsing (n-gram based machine learning and cascaded finite-state parsing), and evaluated them on the basis of a large amount of data. A new representation format has been introduced that allows under-specification of select types of syntactic ambiguity (attachment and subcategorization) even in the absence of a full-fledged chart. Several methods have been discussed for combining the two approaches.</Paragraph>
    <Paragraph position="1"> It has been shown that while combination with the shallow approach can only marginally improve performance of the cascaded parser if ideal disambiguation is assumed, a quite substantial rise is registered in situations closer to the real world where POS tagging is deficient and resolution of attachment and subcategorization ambiguities less than perfect.</Paragraph>
    <Paragraph position="2"> In ongoing work, we look at integrating a statistic context-free parser called BitPar, which was written by Helmut Schmid and achieves .816 F-score on NEGRA. Interestingly, the performance goes up to .9474 F-score when BitPar is combined with the FS parser (upper bound) and .9443 for the lower bound.</Paragraph>
    <Paragraph position="3"> So at least for German, combining parsers seems to be a pretty good idea. Thanks are due to Helmut Schmid and Prof. C. Rohrer for discussions, and to the reviewers for their detailed comments.</Paragraph>
  </Section>
class="xml-element"></Paper>