<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-0706">
  <Title>Exploring Evidence for Shallow Parsinga0</Title>
  <Section position="5" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> Full parsing and shallow parsing are two different strategies for parsing natural languages. While full parsing provides more complete information about a sentence, shallow parsing is more flexible, easier to train and could be targeted for specific, limited subtasks. Several arguments have been used in the past to argue, on an intuitive basis, that (1) shallow parsing is sufficient for a wide range of applications and that (2) shallow parsing could be more reliable than full parsing in handling ill-formed real-world sentences, such as sentences that occur in conversational situations.</Paragraph>
    <Paragraph position="1"> While the former is an experimental issue that is still open, this paper has tried to evaluate experimentally the latter argument. Although the experiments reported here only compare the performance of one full parser and one shallow parser, we believe that these state-of-the-art parsers represent their class quite well. Our results show that on the specific tasks for which we have trained the shallow parser - identifying several kinds of phrases - the shallow parser performs more accurately and more robustly than the full parser. In some sense, these results validate the research in this direction.</Paragraph>
    <Paragraph position="2"> Clearly, there are several directions for future work that this preliminary work suggests. First, in our experiments, the Collins' parser is trained on the Treebank data and tested on the lower quality data. It would be interesting to see what are the results if lower quality data is also used for training. Second, our decision to run the experiments on two different ways of decomposing a sentence into phrases was somewhat arbitrary (although we believe that selecting phrases in a different way would not affect the results). It does reflect, however, the fact that it is not completely clear what kinds of shallow parsing information should one try to extract in real applications. Making progress in the direction of a formal definition of phrases and experimenting with these along the lines of the current study would also be useful.</Paragraph>
    <Paragraph position="3"> Finally, an experimental comparison on several other shallow parsing tasks such as various attachments and relations detection is also an important direction that will enhance this work.</Paragraph>
  </Section>
class="xml-element"></Paper>