<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2328">
  <Title>Multi-level Dialogue Act Tags</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have discussed some issues concerning the design and use of dialogue act tagsets. It is too early to draw rm conclusions from this preliminary study.</Paragraph>
    <Paragraph position="1"> We can note the obvious point that simpli ed smaller tagsets are easier to predict accurately than larger ones. There appear to be non-trivial dependencies between the tags for reasons that are not yet clear.</Paragraph>
    <Paragraph position="2"> We expect the performance of a nal, fully automatic classi er to be substantially higher than the results presented here, owing to the use of more powerful classi ers and, more importantly, larger and richer feature sets. Finally we note that an important point of tagset design has not been addressed empirically here: the question of whether particular distinctions in the tagset are actually useful in our application.</Paragraph>
    <Paragraph position="3"> Future studies will address this point by studying the queries formulated by potential users of meeting processing and retrieval systems.</Paragraph>
  </Section>
class="xml-element"></Paper>