<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1302">
  <Title>Hiroshi Tsujino +</Title>
  <Section position="3" start_page="0" end_page="9" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Many spoken dialogue systems have been developed for various domains, including: flight reservations (Levin et al., 2000; Potamianos and Kuo, 2000; San-Segundo et al., 2000), train travel information (Lamel et al., 1999), and bus information (Komatani et al., 2005b; Raux and Eskenazi, 2004). Since these systems only handle a single domain, users must be aware of the limitations of these domains, which were defined by the system developer. To handle various domains through a single interface, we have developed a multi-domain spoken dialogue system, which is composed of several single-domain systems. The system can handle complicated tasks that contain requests across several domains.</Paragraph>
    <Paragraph position="1"> Multi-domain spoken dialogue systems need to satisfy the following two requirements: (1) extensibility and (2) robustness against speech recognition errors. Many such systems have been developed on the basis of a master-slave architecture, which is composed of a single master module and several domain experts handling each domain.</Paragraph>
    <Paragraph position="2"> This architecture has the advantage that each domain expert can be independently developed, by modifying existing experts or adding new experts into the system. In this architecture, the master module needs to select a domain expert to which response generation and dialogue management for the user's utterance are committed. Hereafter, we will refer to this selecting process domain selection. null The second requirement is robustness against speech recognition errors, which are inevitable in systems that use speech recognition. Therefore, these systems must robustly select domains even when the input may be incorrect due to speech recognition errors.</Paragraph>
    <Paragraph position="3"> We present an architecture for a multi-domain spoken dialogue system that incorporates a new domain selection method that is both extensible and robust against speech recognition errors. Since our system is based on extensible architecture similar to that developed by O'Neill (O'Neill et al., 2004), we can add and modify the domain  experts easily. In order to maintain robustness, domain selection takes into consideration various features concerning context and situations of the dialogues. We also designed a new selection framework that satisfies the extensibility issue by abstracting the transitions between the current and next domains. Specifically, our system selects the next domain based on: (I) the previous domain, (II) the domain in which the speech recognition result can be accepted with the highest recognition score, and (III) other domains. Conventional methods cannot select the correct domain when neither the previous domain nor the speech recognition results for a current utterance are correct. To overcome this drawback, we defined another choice as (III) that enables the system to detect an erroneous situation and thus prevent the dialogue from continuing to be incorrect. We modeled this framework as a classification problem using machine learning, and showed it is effective by performing an experimental evaluation of 2,205 utterances collected from 10 subjects.</Paragraph>
  </Section>
class="xml-element"></Paper>