<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1412">
  <Title>Noun Phrase Generation for Situated Dialogs</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In today's world of mobile, context-aware computing, intelligent software agents are being deployed in a wide variety of domains to aid humans in performing navigation tasks. Examples include hand-held tourist information portals (Johnston et al., 2002) campus tour guides (Yang et al., 1999; Long et al., 1996; Striegnitz et al., 2005), direction-giving avatars for visitors to a building (Cassell et al., 2002; Chou et al., 2005), in-car driving direction systems (Dale et al., 2003; Wahlster et al., 2001), and pedestrian navigation systems (Muller, 2002). These applications present an exciting and challenging new frontier for dialog agents, since attributes of the real-world setting must be combined with other contextual factors for the agent to communicate successfully.</Paragraph>
    <Paragraph position="1"> In the current work, we focus on a scenario in which the system provides incremental directions to a mobile user who is following the instructions as they are produced. Unlike the rigid directions produced by applications like Mapquest,  which describes the entire route from start to finish, this task requires realtime instructions issued while monitoring the user's progress. Instructions are based on dynamic local context variables such as the visibility of and distance to reference points. In referring to items in the setting, human speakers produce a wide variety of noun phrase forms, including descriptions that are headed by a common noun and that employ a definite, indefinite, or demonstrative determiner, one anaphors, and pronouns such as it, this and that. Our goal in the current work is to model that entire space of variation, which makes the task more difficult than the noun phrase generation task defined in many previous studies that simplify the alternatives down to description or pronoun.</Paragraph>
    <Paragraph position="2"> In order to study this process, we developed a task domain in which a human partner is directed through an interior space (a graphically-presented 3D virtual world) to perform a sequence of manipulation tasks. In the first stages of the work, we collected and annotated a corpus of human-human dialogs from this domain. Then, using this data, we trained a decision-tree classifier to utilize context variables such as distance, target object visibility, discourse history, etc., to determine lexical properties of referring expressions to be produced by the generation component of our dialog system.</Paragraph>
  </Section>
class="xml-element"></Paper>