<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2916">
  <Title>Unsupervised Grammar Induction by Distribution and Attachment</Title>
  <Section position="8" start_page="271" end_page="271" type="concl">
    <SectionTitle>
8 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented an incremental grammar induction system that uses heuristics to improve the efficiency of distributional learning. However, in tests over a large corpus, we have shown that it is capable of learning only a small subset of constituent structure. We have analyzed actual constituent-context distributions to explain these limitations. This analysis provides the motivation for a more structured learning method, which incorporates knowledge of verifiable constituent boundaries - the starts and ends of sentences. This improved system performs significantly better, with a 75% increase in recall over distributional methods, and a significant improvement at retrieving structures that are problematic for distributional methods alone.</Paragraph>
  </Section>
class="xml-element"></Paper>