<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3406">
  <Title>Improving &amp;quot;Email Speech Acts&amp;quot; Analysis via N-gram Selection</Title>
  <Section position="3" start_page="35" end_page="36" type="intro">
    <SectionTitle>
2 &amp;quot;Email-Acts&amp;quot; Taxonomy and
Applications
</SectionTitle>
    <Paragraph position="0"> A taxonomy of speech acts applied to email communication (email-acts) is described and motivated in (Cohen et al., 2004). The taxonomy was divided into verbs and nouns, and each email message is represented by one or more verb-noun pairs. For example, an email proposing a meeting and also requesting a project report would have the labels Propose-</Paragraph>
    <Section position="1" start_page="35" end_page="36" type="sub_section">
      <SectionTitle>
Meeting and Request-Data.
</SectionTitle>
      <Paragraph position="0"> The relevant part of the taxonomy is shown in Figure 1. Very briefly, a Request asks the recipient to perform some activity; a Propose message proposes a joint activity (i.e., asks the recipient to perform some activity and commits the sender); a Commit message commits the sender to some future course of action; Data is information, or a pointer to information, delivered to the recipient; and a Meeting is a joint activity that is constrained in time and (usually) space.</Paragraph>
      <Paragraph position="1"> Several possible verbs/nouns were not considered here (such as Refuse, Greet, and Remind), either because they occurred very infrequently in the corpus, or because they did not appear to be important for task-tracking. The most common verbs found in the labeled datasets were Deliver, Request, Commit, and Propose, and the most common nouns were Meeting and deliveredData (abbreviated as dData henceforth). null In our modeling, a single email message may have multiple verbs-nouns pairs.</Paragraph>
      <Paragraph position="2">  ments. Shaded nodes are the ones for which a classifier was constructed.</Paragraph>
      <Paragraph position="3"> Cohen et al. (2004) showed that machine learning algorithms can learn the proposed email-act categories reasonably well. It was also shown that there is an acceptable level of human agreement over the categories. In experiments using different human annotators, Kappa values between 0.72 and 0.85 were obtained. The Kappa statistic (Carletta, 1996) is typically used to measure the human inter-rater agreement. Its values ranges from -1 (complete disagreement) to +1 (perfect agreement) and it is defined as (A-R)/(1-R), where A is the empirical probability of agreement on a category, and R is the probability of agreement for two annotators that  label documents at random (with the empirically observed frequency of each label).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>