<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3405">
  <Title>Shallow Discourse Structure for Action Item Detection</Title>
  <Section position="3" start_page="0" end_page="31" type="intro">
    <SectionTitle>
2 Background
</SectionTitle>
    <Paragraph position="0"> Related Work Corston-Oliver et al. (2004) attempted to identify action items in e-mails, using classifiers trained on annotations of individual sentences within each e-mail. Sentences were annotated with one of a set of &amp;quot;dialogue&amp;quot; act classes; one class Task included any sentence containing items that seemed appropriate to add to an ongoing to-do list. They report good inter-annotator agreement over their general tagging exercise (k &gt; 0.8), although individual figures for the Task class are not given. They then concentrated on Task sentences, establishing a set of predictive features (in which word n-grams emerged as &amp;quot;highly predictive&amp;quot;) and achieved reasonable per-sentence classification performance (with f-scores around 0.6).</Paragraph>
    <Paragraph position="1"> While there are related tags for dialogue act tagging schema - like DAMSL (Core and Allen, 1997), which includes tags such as Action-Directive and Commit, and the ICSI MRDA schema (Shriberg et al., 2004) which includes a commit tag - these classes are too general to allow identification of action items specifically. One comparable attempt in spoken discourse took a flat approach, annotating utterances as action-item-related or not (Gruenstein et al., 2005) over the ICSI and ISL meeting corpora (Janin et al., 2003; Burger et al., 2002). Their inter-annotator agreement was low (k = .36). While this may have been partly due to their methods, it is notable that (Core and Allen, 1997) reported even lower agreement (k = .15) on their Commit dialogue acts. Morgan et al. (forthcoming)thenusedtheseannotationstoattemptauto- null  matic classification, but achieved poor performance (with f-scores around 0.3 at best).</Paragraph>
    <Paragraph position="2"> Action Items Action items typically embody the transfer of group responsibility to an individual.</Paragraph>
    <Paragraph position="3"> This need not be the person who actually performs the action (they might delegate the task to a subordinate), but publicly commits to seeing that the action is carried out; we call this person the owner of the action item. Because this action is a social action that is coordinated by more than one person, its initiation is reinforced by agreement and uptake among the owner and other participants that the action should and will be done. And to distinguish this action from immediate actions that occur during the meeting and from more vague future actions that are still in the planning stage, an action item will be specified as expected to be carried out within a timeframethatbeginsatsomepointafterthemeetingand null extendsnofurtherthanthenot-too-distantfuture. So an action item, as a type of social action, often comprises four components: a task description, a timeframe, an owner, and a round of agreement among the owner and others. The related discourse tends to reflect this, and we attempt to exploit this fact here.</Paragraph>
  </Section>
class="xml-element"></Paper>