<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-0208">
  <Title>Learning Domain-Specific Information Extraction Patterns from the Web</Title>
  <Section position="4" start_page="66" end_page="66" type="intro">
    <SectionTitle>
2 The MUC-4 IE Task and Data
</SectionTitle>
    <Paragraph position="0"> The focus of our research is on the MUC-4 information extraction task (Sundheim, 1992), which is to extract information about terrorist events. The MUC-4 corpus contains 1700 stories, mainly news articles related to Latin American terrorism, and associated answer key templates containing the information that should be extracted from each story.</Paragraph>
    <Paragraph position="1"> We focused our efforts on two of the MUC-4 string slots, which require textual extractions: human targets (victims) and physical targets. The MUC-4 data has proven to be an especially difficult IE task for a variety of reasons, including the fact that the texts are entirely in upper case, roughly 50% of the texts are irrelevant (i.e., they do not describe a relevant terrorist event), and many of the stories that are relevant describe multiple terrorist events that need to be teased apart. The best results reported across all string slots in MUC-4 were in the 50%-70% range for recall and precision (Sundheim, 1992), with most of the MUC-4 systems relying on heavily handengineered components. Chieu et al. (2003) recently developed a fully automatic template generator for the MUC-4 IE task. Their best system produced recall scores of 41%-44% with precision scores of 49%-51% on the TST3 and TST4 test sets.</Paragraph>
  </Section>
class="xml-element"></Paper>