<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2410">
  <Title>Thesauruses for Prepositional Phrase Attachment</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> Our results show that the similarity-based smoothing of frequency estimates significantly improves an already respectable probabilistic PP attachment model. However our hypothesis that a task-specific thesaurus would out-perform a generic thesaurus was not borne out by our experiments. The neighbours provided by the specialist thesaurus are not as informative as those supplied by the generic thesauruses. Of course, this negative result is naturally good news for developers of generic thesauruses.</Paragraph>
    <Paragraph position="1"> We described ways of finding and scoring distributionally similar PPs. A significant number of errors in the final model can be traced to the way individual words in the PP are replaced without regard to the wider context, producing neighbouring PPs that have conflicting attachment preferences. The specialist thesaurus was not able to overcome this problem. A second finding is that distributional similarity scores provided by all thesauruses weight dissimilar neighbours too highly, and more aggressive weighting schemes are better for smoothing.</Paragraph>
    <Paragraph position="2"> Our aim is to apply similarity-based smoothing with both generic and specialist thesauruses to other areas in lexicalised parse selection, particularly other overtly lexical problems such as noun-noun modifiers and conjunction scope. Lexical information has a lot of promise for parse selection in theory, but there are practical problems such as sparse data and genre effects (Gildea, 2001). Appropriately trained thesauruses and similarity-based techniques should help to alleviate both problems.</Paragraph>
  </Section>
class="xml-element"></Paper>