<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1303">
  <Title>Building Effective Question Answering Characters</Title>
  <Section position="9" start_page="23" end_page="24" type="concl">
    <SectionTitle>
6 Conclusions and future work
</SectionTitle>
    <Paragraph position="0"> In this paper we presented a method for efficient construction of conversational virtual characters.</Paragraph>
    <Paragraph position="1"> These characters accept spoken input from a user, convert it to text, and select the appropriate response using statistical language modeling techniques from cross-lingual information retrieval.</Paragraph>
    <Paragraph position="2"> We showed that in this domain the performance of our answer selection approach significantly exceeds the performance of a state of the art text classification method. We also showed that our technique is very robust to the quality of the input and can be effectively used with existing speech recognition technology.</Paragraph>
    <Paragraph position="3"> Preliminary failure analysis indicates a few directions for improving the system's quality. First, we should continue collecting more training data and extending the question sets.</Paragraph>
    <Paragraph position="4"> Second, we could have the system generate a confidence score for its classification decisions.</Paragraph>
    <Paragraph position="5"> Then the answers with a low confidence score can be replaced with an answer that prompts the user to rephrase her question. The system would then  use the original and the rephrased version to repeat the answer selection process.</Paragraph>
    <Paragraph position="6"> Finally, we observed that a notable percent of misclassifications results from the user asking a question that has a strong context dependency on the previous answer or question. We are presently looking into incorporating this context information into the answer selection process.</Paragraph>
  </Section>
class="xml-element"></Paper>