<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1106">
  <Title>References</Title>
  <Section position="6" start_page="55" end_page="55" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have shown that conceptual case frames can be generated automatically using unannotated text as input, coupled with a few hours of manual review. Our results for the terrorism domain show that the case frames achieve similar recall levels as the extraction patterns, but with substantially fewer false hits. Our results are not directly comparable to the MUC-4 results because the MUC-4 systems contained additional components, such as domain-specific discourse analyzers that resolved coreferent noun phrases, merged event descriptions, and illtered out irrelevant information. The work presented here only addresses the initial stage of information extraction. However, in previous work we showed that AutoSlog-TS achieved performance comparable to AutoSlog (Riloff, 1996b), which performed very well in the MUC-4 evaluation (Lehnert et al., 1992b). Since the conceptual case frames achieved comparable recall and higher precision than AutoSlog-TS' extraction patterns, our results suggest that the case frames performed well relative to previous work on this domain.</Paragraph>
    <Paragraph position="1"> Several other systems learn extraction patterns that can also be viewed as conceptual case frames with selectional restrictions (e.g., PALKA (Kim and Moldovan, 1993) and CRYSTAL (Soderland et al., 1995)). The case frames learned by our system are not necessarily more powerful then those generated by other systems. The advantage of our approach is that it requires no special training resources. Our technique requires only preclassified training texts and a few hours of manual filtering to build the intermediate dictionaries. Given preclassified texts, it is possible to build a dictionary of conceptual case frames for a new domain in one day.</Paragraph>
    <Paragraph position="2"> Another advantage of our approach is its highly empirical nature; a corpus often reveals important patterns in a domain that are not necessarily intuitive to people. By using corpus-based methods to generate all of the intermediate dictionaries and the final case frame structures, the most important words, role assignments, and semantic prefe:rences are less likely to be missed. Our empirical approach aims to exploit the text corpus to automatically acquire the syntactic and semantic role assignments that are necessary to achieve good performalace in the domain.</Paragraph>
  </Section>
class="xml-element"></Paper>