<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-1041">
  <Title>Answer Extraction</Title>
  <Section position="5" start_page="299" end_page="300" type="concl">
    <SectionTitle>
4 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We have described a system that handles arbitrary questions, producing a candidate list of answers ranked by their plausibility. Evaluation on the TREC question-answering track showed that the correct answer to queries appeared in the top five answers 46% of the time, with a mean score of 0.356.</Paragraph>
    <Paragraph position="1"> The average length of answers produced by the system was 10.5 bytes.</Paragraph>
    <Paragraph position="2"> 4The sole exception was a query misclassification caused by a parse failure---miraculously, the correct answer made it to rank five despite being of the &amp;quot;wrong&amp;quot; type.  There are several possible areas for future work.</Paragraph>
    <Paragraph position="3"> There may be potential for improved performance through more sophisticated use of NLP techniques.</Paragraph>
    <Paragraph position="4"> In particular, the syntactic context in which a particular entity appears may provide important information, but it is not currently used by the system. Another area of future work is to extend the entity-extraction component of the system to handle arbitrary types (mountain ranges, films etc.). The error analysis in section 3.2.2 showed that these question types cause particular difficulties for the system.</Paragraph>
    <Paragraph position="5"> The system is largely hand-built. It is likely that as more features are added a trainable statistical or machine learning approach to the problem will become increasingly desirable. This entails developing a training set of question-answer pairs, raising the question of how a relatively large corpus of questions can be gathered and annotated.</Paragraph>
  </Section>
class="xml-element"></Paper>