<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0213">
  <Title>Dialogue Act Recognition with Bayesian Networks for Dutch Dialogues</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In several papers (Nijholt, 2000; Luin et al., 2001; Nijholt et al., 2001) we reported on our virtual music centre - the VMC - a virtual environment inhabited by (embodied) agents and on multi-modal interaction between human users and these agents. Of these agents Karin is an embodied agent users can ask for information about theatre performances (see Figure 1).</Paragraph>
    <Paragraph position="1"> A second agent is the navigation agent. Navigation is a) way finding - the user knows where he wants to go but doesn't know how to go there; or b) exploring the environment - the user walks through the environment to obtain an overview of the building and the objects, locations, rooms that are in it. Related to these navigation tasks the navigation assistant has the task to assist the visitor in a) explaining how to go from his current loca-tion to a location he is looking for and b) to give the agent information about objects, and locations in the environment. The navigation agent is not present as an avatar in the environment. The user sees the environment from a first person perspective and interacts with the agents by means of a Dutch dialogue. The user has two views of the environment: a) a first person view of the visible part of the 3D virtual theatre and b) an abstract 2D map of the floor of the building the user is visiting. This map is shown in a separate window. In a multi-modal interaction the user can point at locations or objects on the 2D map and either ask information about that object or loca-tion or he can ask the assistant to bring him to the  An important part of our dialogue systems for natural language interaction with agents is the module for recognition of the dialogue acts performed by the human user (visitor). This paper discusses the construction of and experiments with Bayesian networks as implementation of this module.</Paragraph>
    <Paragraph position="2"> Various other work has been presented on using statistical techniques for dialogue act classification (Andernach, 1996; Stolcke et al., 2000), and even some first efforts on using Bayesian networks for this task (Pulman, 1996; Keizer, 2001). Other work on using Bayesian networks in dialogue systems aims more at interaction and user modelling (Paek and Horvitz, 2000) and does not specifically involve linguistic aspects.</Paragraph>
    <Paragraph position="3"> The paper is organised as follows. Section 2 Philadelphia, July 2002, pp. 88-94. Association for Computational Linguistics. Proceedings of the Third SIGdial Workshop on Discourse and Dialogue, provides some necessary and general background about the use of Bayesian networks for speech act recognition. In Section 3 we discuss experiments with a Bayesian network for dialogue act classification based on a dialogue corpus for the Karin agent. In Section 4 we discuss our current experiments with a network for the navigation dialogue system that was automatically created from a small corpus. Section 5 reflects on our findings and presents plans for the near future.</Paragraph>
  </Section>
class="xml-element"></Paper>