<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2040">
  <Title>Sentence Planning for Realtime Navigational Instructions</Title>
  <Section position="5" start_page="159" end_page="159" type="concl">
    <SectionTitle>
4 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We describe an experiment in content planning for spoken dialog agents that provide navigation instructions. Navigation requires the system and the user to achieve joint reference to objects in the environment. To accomplish this goal human directiongivers judge whether their partner is in an appropriate spatial con guration to comprehend a reference spoken to an object in the scene. If not, one strategy for accomplishing the communicative goal is to steer their partner into a position from which the object is easier to describe.</Paragraph>
    <Paragraph position="1"> The algorithm we developed in this study, which takes into account spatial context features replicates our human subject's decision to produce a description with 86%, compared to a 70% baseline based on the visibility of the object. Although the spatial details will vary for other spoken dialog domains, the process developed in this study for producing description dialog moves only at the appropriate times should be relevant for spoken dialog agents operating in other navigation domains.</Paragraph>
    <Paragraph position="2"> Building dialog agents for situated tasks provides a wealth of opportunity to study the interaction between context and linguistic behavior. In the future, the generation procedure for our interactive agent will be further developed in areas such as spatial descriptions and surface realization. We also plan to investigate whether different object types in the domain require differential processing, as prior work on spatial semantics would suggest.</Paragraph>
  </Section>
class="xml-element"></Paper>