<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1070">
  <Title>Inducing Information Extraction Systems for New Languages via Cross-Language Projection</Title>
  <Section position="9" start_page="7" end_page="7" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have used IE systems for English to automatically derive IE systems for a second language. Even with the quality of MT available today, our results demonstrate that we can exploit translation tools to transfer information extraction expertise from one language to another. Given an IE system for a source language, an MT system that can translate between the source and target languages, and a word alignment algorithm, our approach allows a user to create a functionally comparable IE system for the target language with very little human effort. Our experiments demonstrated that the new IE system can achieve roughly the same level of performance as the source-language IE system. French and English are relatively close languages, however, so how well these techniques will work for more distant language pairs is still an open question.</Paragraph>
    <Paragraph position="1"> Additional performance benefits could be achieved in two ways: (1) put more effort into obtaining better resources for English, or (2) implement (minor) specializations per language.</Paragraph>
    <Paragraph position="2"> While it is expensive to advance the state of the art in English IE or to buy annotated data for a new domain, these additions will improve performance not only in English but for other languages as well. On the other hand, with minimal effort (hours) it is possible to custom-train a system such as Autoslog/Sundance to work relatively well on noisy MT-English, providing a substantial performance boost for the IE system learned for the target language, and further gains are achieved via voting-based classifier combination.</Paragraph>
  </Section>
class="xml-element"></Paper>