<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2004">
  <Title>The Effect of Corpus Size in Combining Supervised and Unsupervised Training for Disambiguation</Title>
  <Section position="9" start_page="30" end_page="31" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> Previous work on specific types of ambiguities (like RC and PP) has not addressed the integration of specific resolution algorithms into a generic statistical parser. In this paper, we have shown for two types of ambiguities, relative clause and prepositional phrase attachment ambiguity, that integration into a statistical parser is possible and that the com6Strictly speaking, our experiments were not completely unsupervised since the default value and the most frequent attachment were determined based on the development set.</Paragraph>
    <Paragraph position="1"> 7We attempted to recreate Siddharthan's training and test sets, but were not able to based on the description in the paper and email communication with the author.</Paragraph>
    <Paragraph position="2">  bined system performs better than either component by itself. However, for PP attachment this only holds for small training set sizes. For large training sets, we could only show an improvement for RC attachment.</Paragraph>
    <Paragraph position="3"> Surprisingly, we only found a small effect of the size of the unlabeled corpus on disambiguation performance due to the noisiness of statistics extracted from raw text. Once the unlabeled corpus has reached a certain size (510 million words in our experiments) combined performance does not increase further.</Paragraph>
    <Paragraph position="4"> The results in this paper demonstrate that the baseline of a state-of-the-art lexicalized parser for specific disambiguation problems like RC and PP is quite high compared to recent results for stand-alone PP disambiguation. For example, (Toutanova et al., 2004) achieve a performance of 87.6% for a training set of about 85% of WSJ. That number is not that far from the 82.8% achieved by Collins' parser in our experiments when trained on 50% of WSJ. Some of the supervised approaches to PP attachment may have to be reevaluated in light of this good performance of generic parsers.</Paragraph>
  </Section>
class="xml-element"></Paper>