Information States and Dialogue Move Engines

Peter Bohlin, Robin Cooper, Elisabet Engdahl, Sta an Larsson Department of linguisti s G oteborg University Box 200, Humanisten, SE-405 30 Goteborg, SWEDEN fpeb, ooper,engdahl,slg ling.gu.se Abstra t We explore the notion of information state in relation to dialogue systems, and in parti ular to the part of a dialogue system we all the dialogue move engine. We use a framework for experimenting with information states and dialogue move engines, and show how an experimental dialogue system urrently being developed in G oteborg within the framework an be provided with rules to handle a ommodation of questions and plans in dialogue. 1 Introdu tion We use the term information state to mean, roughly, the information stored internally by an agent, in this ase a dialogue system. A dialogue move engine updates the information state on the basis of observed dialogue moves and sele ts appropriate moves to be performed. In this paper we use a formal representation of dialogue information states that has been developed in the TRINDI1, SDS2 and INDI3 proje ts4. The stru ture of this paper is as follows: First, we give a brief des ription of a general dialogue system ar hite ture whi h an be used for experimenting with di erent kinds of information states and dialogue move engines. We explain the distin tion between stati and dynami information state, and dis uss how rules formulated in terms of onditions and operations on in1TRINDI (Task Oriented Instru tional Dialogue), EC Proje t LE4-8314, www.ling.gu.se/resear h/proje ts/trindi/ 2SDS (Swedish Dialogue Systems), NUTEK/HSFR Language Te hnology Proje t F1472/1997, http://www.ida.liu.se/ nlplab/sds/ 3INDI (Information Ex hange in Dialogue), Riksbankens Jubileumsfond 1997-0134. 4We will illustrate our dis ussion from a Swedish dialogue in the travel booking domain that has been olle ted by the University of Lund as part of the SDS proje t. We quote the trans ription done in G oteborg as part of the same proje t. formation states an be used to (1) update information states based on observed dialogue moves and (2) sele t dialogue moves based on the urrent information state. We present a parti ular notion of dynami information state based on Ginzburg's theory of Questions Under Dis ussion (QUD) [Ginzburg, 1996a; 1996b; 1998℄. An experimental dialogue system whi h uses this notion of information state is presented. We then look at the role of a ommodation in information state transitions and point to examples of two kinds of a ommodation: a ommodation of questions under dis ussion and of dialogue plan. Finally, we argue that a ommodation should be asso iated with update rules, or ta it moves (not asso iated with an utteran e), and show how the implementation of these rules yields improved behaviour in the experimental dialogue system. 2 General ar hite ture The general ar hite ture we are assuming is shown in (1). (1)