<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3405">
  <Title>Shallow Discourse Structure for Action Item Detection</Title>
  <Section position="7" start_page="33" end_page="33" type="concl">
    <SectionTitle>
6 Discussion and Future Work
</SectionTitle>
    <Paragraph position="0"> So accounting for the structure of action items appears essential to detecting them in spoken discourse. Otherwise, classification accuracy is limited. We believe that accuracy can be improved, and the detected utterances can be used to provide the properties of the action item itself. An interesting question is how and whether the structure we use here relates to discourse structure in more general use. If a relation exists, this would shed light on the decision-making process we are attempting to (begin to) model, and might allow us to use other (more plentiful) annotated data.</Paragraph>
    <Paragraph position="1"> Our future efforts focus on annotating more meetings to obtain large training and testing sets. We also wish to examine performance when working from speech recognition hypotheses (as opposed to the human transcripts used here), and the best way to incorporate multiple hypotheses (either as n-best lists or word confusion networks). We are actively investigatingalternativeapproachestosub-classifiercom- null bination: better performance (and a more robust and trainable overall system) might be obtained by using a Bayesian network, or a maximum entropy classifier as used by (Klein et al., 2002). Finally, we are developing an interface to a new large-vocabulary version of the Gemini parser (Dowding et al., 1993) which will allow us to use semantic parse informationasfeaturesintheindividualsub-classclassifiers, null and also to extract entity and event representations from the classified utterances for automatic addition of entries to calendars and to-do lists.</Paragraph>
  </Section>
class="xml-element"></Paper>