<?xml version="1.0" standalone="yes"?>
<Paper uid="P02-1055">
  <Title>Shallow parsing on the basis of words only: A case study</Title>
  <Section position="7" start_page="35" end_page="35" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> POS are normally considered useful information in shallow and full parsing. Our learning curve experiments show that: a84 The relative merit of words versus POS as input for the combined chunking and function-tagging task depends on the amount of training data available.</Paragraph>
    <Paragraph position="1"> a84 The absolute performance of words depends on the treatment of rare words. The additional use of word form information (attenuation) improves performance.</Paragraph>
    <Paragraph position="2"> a84 The addition of POS also improves performance. In this and the previous case, the effect becomes smaller with more training data.</Paragraph>
    <Paragraph position="3"> Experiments with the maximal training set size show that: a84 Addition of POS maximally yields an improvement of 1.7 points on this data.</Paragraph>
    <Paragraph position="4"> a84 With realistic POS the improvement is much smaller.</Paragraph>
    <Paragraph position="5"> Preliminary analysis shows that the improvement by realistic POS seems to be caused mainly by a superior use of word form information by the POS tagger. We therefore plan to experiment with a POS tagger and an attenuated words variant that use exactly the same word form information. In addition we also want to pursue using the combined chunker and grammatical function tagger described here as a first step towards grammatical relation assignment.</Paragraph>
  </Section>
class="xml-element"></Paper>