<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-3001">
  <Title>An Information-State Approach to Collaborative Reference</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
2 Overview and Related Work
</SectionTitle>
    <Paragraph position="0"> Our demonstration system plays a referential communication game, much like the one that pairs of human subjects play in the experiments of Clark and Wilkes-Gibbs (1986). We describe each episode in this game as an activity involving the coordinated action of two participants: a director D who knows the referent R of a target variable T and a matcher M whose task is to identify R. Our system can play either role, D or M, using virtual objects in a graphical display as candidate targets and distractors, and using text as its input and output. Our system uses the same task knowledge and the same grammar whichever role it plays. Of course, the system also draws on private knowledge to decide how best to carry out its role; for now it describes objects using the domain-specific iteration proposed by Dale and Reiter (1995). The knowledge we have formalized is targeted to a proof-of-concept implementation, but we see no methodological obstacle in adding to the  system's resources.</Paragraph>
    <Paragraph position="1"> We exemplify what our system does in (1).</Paragraph>
    <Paragraph position="2">  (1) a. S: This one is a square.</Paragraph>
    <Paragraph position="3"> b. U: Um-hm...</Paragraph>
    <Paragraph position="4"> c. S: It's light brown.</Paragraph>
    <Paragraph position="5"> d. U: You mean like tan? e. S: Yeah.</Paragraph>
    <Paragraph position="6"> f. S: It's solid.</Paragraph>
    <Paragraph position="7"> g. U: Got it.</Paragraph>
    <Paragraph position="8">  The system (S) and user (U) exchange seven utterances in the course of identifying a tan solid square. We achieve this interaction using the information-state approach to dialogue system design (Larsson and Traum, 2000). This approach describes dialogue as a coordinated effort to maintain an agreed record of the state of the conversation. Our model contrasts with traditional plan-based models, as exemplified by Heeman and Hirst's model of goals and beliefs in collaborative reference (1995). Our approach abstracts away from such details of individuals' mental states and cognitive processes, for principled reasons (Stone, 2004a). We are able to capture these details implicitly in the dynamics of conversation, whereas plan-based models must represent them explicitly. Our representations are simpler than Heeman and Hirst's but support more flexible dialogue. For example, their approach to (1) would have interlocutors coordinating on goals and beliefs about a syntactic representation for the tan solid square; for us, this description and the interlocutors' commitment to it are abstract results of the underlying collaborative activity.</Paragraph>
    <Paragraph position="9"> Another important antecedent to our work is Purver's (2004) characterization of clarification of names for objects and properties. We extend this work to develop a treatment of referential descriptive clarification. When we describe things, our descriptions grow incrementally and can specify as much detail as needed. Clarification becomes correspondingly cumulative and open-ended. Our revised information state includes a model of cumulative and open-ended collaborative activity, similar to that advocated by Rich et al. (2001). We also benefit from a reversible goal-directed perspective on descriptive language (Stone et al., 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>