<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1017">
  <Title>Exploiting Strong Syntactic Heuristics and Co-Training to Learn Semantic Lexicons</Title>
  <Section position="5" start_page="5" end_page="5" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented a method for learning semantic lexicons that uses strong syntactic heuristics in a bootstrapping algorithm. We exploited three types of syntactic structures (appositives, compound NPs,  After co-training finished, we also added terms to the lexicon that were hypothesized by an individual classifier with frequency &lt; if they had not previously been labeled.</Paragraph>
    <Paragraph position="1"> and ISA clauses) in combination with heuristics to identify instances of these structures that contain both a proper and general noun phrase. Each syntactic structure generated many lexicon entries, in most cases with high accuracy. We also combined the three classifiers using co-training. The co-training model increased the number of learned lexicon entries, while maintaining nearly the same level of accuracy. One limitation of this work is that it can only learn semantic categories that are commonly found as proper nouns and general nouns.</Paragraph>
    <Paragraph position="2"> This research illustrates that common syntactic structures can be combined with heuristics to identify specific semantic relationships. So far we have experimented with three structures and one type of heuristic (proper NP/general NP collocations), but we believe that this approach holds promise for other semantic learning tasks as well. In future work, we hope to investigate other types of syntactic structures that may be used to identify semantically related terms, and other types of heuristics that can reveal specific semantic relationships.</Paragraph>
  </Section>
class="xml-element"></Paper>