<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1306">
  <Title>Multidimensional Dialogue Management</Title>
  <Section position="3" start_page="0" end_page="37" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> During (task-oriented) dialogues, the participants have to deal with many different aspects of communication simultaneously. Besides some underlying task that may be performed through the dialogue, there are also various aspects of managing the communicative process itself, including dealing with social obligations. Therefore, speakers often use utterances that are multifunctional.</Paragraph>
    <Paragraph position="1"> We will present an approach to dialogue management that accounts for the generation of multi-functional utterances. The approach is based on a dialogue theory involving a multidimensional dialogue act taxonomy and associated context model.</Paragraph>
    <Paragraph position="2"> In this theory, called Dynamic Interpretation Theory (DIT) (Bunt, 1996; Bunt, 2000a), a dialogue is modelled as a sequence of (sets of) dialogue acts operating on the Information State of each of the participants. The dialogue acts are organised in a taxonomy that is multidimensional, i.e., each utterance may involve dialogue acts of at most one type from each dimension. The taxonomy has dimensions for aspects like feedback, interactionmanagement, social obligations management and managing the underlying task.</Paragraph>
    <Paragraph position="3"> In a dialogue system developed according to the principles of DIT, the information state is represented through a context model, containing all information considered relevant for interpreting user utterances an generating system utterances in terms of dialogue acts. Hence, given the multidimensionality of the taxonomy, the input interpretation components of the system result in several dialogue acts for each utterance, at most one from each of the dimensions. Using these recognised user dialogue acts, the context model is updated.</Paragraph>
    <Paragraph position="4"> On the other hand, the ultimate task for a dialogue manager component of a dialogue system is deciding which dialogue acts to generate. So, again with the multidimensional organisation of the taxonomy in mind, we argue for a multi-agent approach, in which the dialogue act generation task is divided over several agents that operate in parallel on the context model, each agent being dedicated to the generation of dialogue acts from one particular dimension in the taxonomy. This leads to the design of a number of so-called Di- null alogue Act Agents, including e.g. a task-oriented agent, two feedback agents and an agent dealing with social obligations management.</Paragraph>
    <Paragraph position="5"> The multi-agent approach to dialogue management itself is not new: JASPIS (Turunen and Hakulinen, 2000; Salonen et al., 2004) is a multi-agent framework for dialogue systems which allows for implementations of several agents for the same tasks, varying from input interpretation and output presentation to dialogue management. Depending on the situation, the agent that is most appropriate for a given task is selected in a process involving several so-called 'evaluators'. In JASPIS the multi-agent approach is aimed at flexibility and adaptiveness, while our approach focuses more on supporting multidimensionality in communication.</Paragraph>
    <Paragraph position="6"> In a very general sense, our dialogue managementapproachfollowsaninformationstateupdate null approach similar to the dialogue managers that are developed within the TRINDI framework (Larsson and Traum, 2000). For example, Matheson et al. (2000) describe the implementation of a dialogue management system focusing in the concepts of grounding and discourse obligations.</Paragraph>
    <Paragraph position="7"> An approach to dialogue management which identifies several simultaneous processes in the generation of system utterances, is described in (Stent, 2002). In this approach, which is implemented in the TRIPS dialogue system, dialogue contributions are generated through three core components operating independently and concurrently, using a system of conversation acts organised in several levels (Traum and Hinkelman, 1992).</Paragraph>
    <Paragraph position="8"> Although there are apparent similarities between our approach and that of the TRINDI based dialogue managers and the TRIPS system, there are clear differences as well, which for an important part stem from the system of dialogue acts used and the way the information state is organised. More particularly, the way in which mechanisms for generating dialogue acts along multiple dimensions are modelled and implemented by means of multiple agents, differs from existing approaches. null This paper is organised as follows. First we explain the closely connected DIT notions of dialogue act and information state, and the multi-dimensional dialogue act taxonomy and context model (Sections 2 and 3). We then introduce the multi-agent approach to dialogue management (Section 4) and illustrate it by a description of the current implementation (Section 4.1). This implementation is carried out in the PARADIME project (PARallel Agent-based DIalogue Management Engine), which is part of the multiproject IMIX (Interactive Multimodal Information Extraction). The PARADIME dialogue manager is integrated into an interactive question-answering system that is developed in a collaboration between several projects participating in IMIX. The paper ends with conclusions and directions for future research (Section 5).</Paragraph>
  </Section>
class="xml-element"></Paper>