<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0201">
  <Title>Marineau Heather Hite-Mitchell</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> One of the objectives of this work was to see how well a classifier could perform with a minimum of resources.</Paragraph>
    <Paragraph position="1"> Using no context and only surface features, the classifier performed with an average weighted F-measure of .98 on real world data.</Paragraph>
    <Paragraph position="2"> However, the question remains how performance will fare as rare questions become more frequent. Scaffolding student questions has become a hot topic recently (Graesser et al. 2003). In a system that greatly promotes question-asking, the weighted average of .97 will tend to drift closer to the unweighted average of .54. Thus there is clearly more work to be done.</Paragraph>
    <Paragraph position="3"> Future directions include using bootstrapping methods and statistical techniques on tutoring corpora and using context to disambiguate question classification.</Paragraph>
  </Section>
class="xml-element"></Paper>