Training a BN-based user model for dialogue simulation with missing data

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 598–604, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Training a BN-based user model for dialogue simulation with missing data St´ephane Rossignol†,‡, Olivier Pietquin†,‡, Michel Ianotto†,‡ †SUPELEC - IMS Research Group 2 rue ´Edouard Belin, Metz France ‡UMI 2958 (GeorgiaTech - CNRS) {forename.surname}@supelec.fr Abstract The design of a Spoken Dialogue System (SDS) is a long, iterative and costly pro- cess. Especially, it requires test phases on actual users either for assessment of per- formance or optimization. The number of test phases should be minimized, yet with- out degrading the final performance of the system. For these reasons, there has been an increasing interest for dialogue simula- tion during the last decade. Dialogue sim- ulation requires simulating the behavior of users and therefore requires user model- ing. User simulation is often done by sta- tistical systems that have to be tuned or trained on data. Yet data are generally in- complete with regard to the necessary in- formation for simulating the user decision making process. For example, the internal knowledge the user builds along the con- versation about the information exchanged while interacting is difficult to annotate. In this contribution, we propose the use of a previously developed user simulation system based on Bayesian Networks (BN) and the training of this model using al- gorithms dealing with missing data. Ex- periments show that this training method increases the simulation performance in terms of similarity with real dialogues. 1 Introduction The design of a Spoken Dialogue System (SDS) is a long, iterative and costly process. Although sev- eral attempts exist to simplify this design such as the VoiceXML language (W3C, 2008), graphical interfaces (McTear, 1998) or machine-learning- based methods (Pietquin and Dutoit, 2003), it re- mains an expert job. Especially, it requires test phases on actual users either for assessment of per- formance (Eckert et al., 1997; L´opez-C´ozar et al., 2006) or strategy optimization by means of rein- forcement learning (Levin et al., 1997; Pietquin and Dutoit, 2006a). The number of test phases should be minimized, yet without degrading the final performance of the system. One solution to this problem is the use of Wizard-of-Oz meth- ods (Kelley, 1984; Rieser, 2008). Although this doesn’t require a real implementation of the di- alogue system to be tested, this is still time and money consuming. For these reasons, there has been an increasing interest for dialogue simula- tion during the last decade (Eckert et al., 1997; Pietquin and Dutoit, 2006a; Schatzmann et al., 2006; L´opez-C´ozar et al., 2006). Dialogue simula- tion requires simulating the behavior of users and therefore requires user modeling as well as error modelling (Pietquin and Dutoit, 2006b; Schatz- mann et al., 2007b). Most often, dialogue simu- lation takes place at the intention level (Eckert et al., 1997; Pietquin and Dutoit, 2006a; Schatzmann et al., 2007c) but can take place at the speech sig- nal level (L´opez-C´ozar et al., 2006). This paper focuses on the former solution …

Content truncated to 500 words