<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3404">
  <Title>You Are What You Say: Using Meeting Participants' Speech to Detect their Roles and Expertise</Title>
  <Section position="3" start_page="23" end_page="24" type="intro">
    <SectionTitle>
2 The Y2 Meeting Scenario Data
</SectionTitle>
    <Paragraph position="0"> Our research work is part of the Cognitive Assistant that Learns and Organizes project (CALO, 2003). A goal of this project is to create an artificial assistant that can understand meetings and use this understanding to assist meeting participants during and after meetings. Towards this goal, data is being collected by creating a rich multimodal record of meetings (e.g. (Banerjee et al., 2004)). While a large part of this data consists of natural meetings (that would have taken place even if they weren't being recorded), a small subset of this data is &amp;quot;scenario driven&amp;quot; - the Y2 Scenario Data.</Paragraph>
    <Section position="1" start_page="23" end_page="24" type="sub_section">
      <SectionTitle>
Meeting # Typical scenario
</SectionTitle>
      <Paragraph position="0"> The Y2 Scenario Data consists of meetings between groups of 3 or 4 participants. Each group participated in a sequence of up to 5 meetings. Each sequence had an overall scenario - the purchasing of computing hardware and the allocation of office space for three newly hired employees. Participants were told to assume that the meetings in the sequence were being held one week apart, and that between any two meetings &amp;quot;progress&amp;quot; was made on the action items decided at each meeting. Participants were given latitude to come up with their own stories of what &amp;quot;progress&amp;quot; was made between meetings. At each meeting, participants were asked to review progress since the last meeting and make changes to their decisions if necessary. Additionally, an extra topic was introduced at each meeting, as shown in table 1.</Paragraph>
      <Paragraph position="1"> In each group of participants, one participant played the role of the manager who has control over the funds and makes the final decisions on the purchases. The remaining 2 or 3 participants played the roles of either the hardware acquisition expert or the building facilities expert. The role of the hardware expert was to make recommendations on the buying of computers and printers, and to actually make the purchases once a decision was made to do so. Similarly the role of the building expert was to make recommendations on which rooms were available to fit the new employees into. Despite this role assign- null ment, all participants were expected to contribute to discussions on all topics.</Paragraph>
      <Paragraph position="2"> To make the meetings as natural as possible, the participants were given control over the evolution of the story, and were also encouraged to create conflicts between the manager's demands and the advice that the experts gave him. For example, managers sometimes requested that all three employees be put in a single office, but the facilities expert announced that no 3 person room was available, unless the manager was agreeable to pay extra for them. These conflicts led to extended negotiations between the participants. To promote fluency, participants were instructed to use their knowledge of existing facilities and equipment instead of inventing a completely fictitious set of details (such as room numbers).</Paragraph>
      <Paragraph position="3"> The data we use in this paper consists of 8 sequences recorded at Carnegie Mellon University and at SRI International between 2004 and 2005. One of these sequences has 4 meetings, the remaining have 5 meetings each, for a total of 39 meetings. 4 of these sequences had a total of 3 participants each; the remaining 4 sequences had a total of 4 participants each. On average each meeting was 15 minutes long. We partitioned this data into two roughly equal sets, the training set containing 4 meeting sequences, and the test set containing the remaining 4 sets. Although a few participants participated in multiple meetings, there was no overlap of participants between the training and the test set.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>