<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2012">
  <Title>Answer Validation by Keyword Association</Title>
  <Section position="12" start_page="5" end_page="5" type="relat">
    <SectionTitle>
7 Related Work
</SectionTitle>
    <Paragraph position="0"> Kwok et al. (2001) proposed the first automated question-answering system which uses the web.</Paragraph>
    <Paragraph position="1"> First, it collects documents that are related to the question sentence using google and picks answer candidates up from them. Second, it selects an answer based on the frequency of candidates which appear near the keywords.</Paragraph>
    <Paragraph position="2"> In the method proposed by Brill et al. (2002), answer candidates are picked up from the summary pages returned by a search engine. Then, each answer candidate is validated by searching for relevant documents in the TREC QA document collection. Both methods do not consider the number of hits returned by the search engine. null Magnini et al. (2002) proposed an answer validation method which uses the number of search engine hits. They formulate search engine queries using AltaVista's OR and NEAR operators. Major difference between the method of Magnini et al. (2002) and ours is in keyword selection. In the method of Magnini et al. (2002), the initial keywords are content words extracted from a question sentence. If the hits of keywords is less than a threshold, the least important key-word is removed. This procedure is repeated until the hits of the keywords is over the threshold. On the other hand, in our method, keywords are selected so that the strength of the association between the keyword and an answer candidate is maximized. Intuitively, our method of keyword selection is more natural than that of Magnini et al. (2002), since it considers both the question sentence and an answer candidate.</Paragraph>
    <Paragraph position="3"> As for measures for scoring answer candidates, Magnini et al. (2002) proposed three measures, out of which &amp;quot;Corrected Conditional Probability&amp;quot; performs best. In our implementation, the performance of &amp;quot;Corrected Conditional Probability&amp;quot; is about 5% lower than our best result.</Paragraph>
  </Section>
class="xml-element"></Paper>