Information-Seeking Chat: Dialogue Management by Topic Structure Manfred Stede and David Schlangen University of Potsdam Department of Linguistics Applied Computational Linguistics P.O. Box 601553 D-14415 Potsdam — Germany {stede|das}@ling.uni-potsdam.de Abstract In this paper we describe the dialogue sub-genre “information-seeking chat”, which is distinguished from other kinds of information-seeking dialogue (e.g. travel information) by its more ex- ploratory and less (single) task-oriented nature. We present an approach to mod- elling this kind of dialogue, based on the notion of weighted topic structures —a single data structure that represents both the domain knowledge and the dialogue history, and we sketch an implementa- tion of this approach in a typed dialogue system. 1 Introduction Both theoretical analyses of dialogue and imple- mented dialogue systems have so far mostly fo- cused on two main dialogue genres: strictly task- oriented dialogue (as in the travel agent domain, call routing applications, or collaborative problem solving domains), or tutorial dialogue. In this pa- per we describe another type of dialogue, which we call “information-seeking chat”. This genre is distinguished by its more exploratory and less task-oriented nature, while still being more struc- tured than general free conversation. Our thesis is that this kind of dialogue can be modelled with a simple taxonomy of dialogue moves and a dialogue management (DM) strategy based on topic structure, where the main task of the dialogue manager is to guide the user through the pre-defined topic map. This topic map is a declarative domain model (similar to an ontology) that serves both as a representation of the domain knowledge and as a repository for the discourse history. (The model represents the discourse his- tory insofar as during the course of the dialogue it is annotated with information about which top- ics have been broached or have been exhausted.) Moreover, it is the only discourse planning de- vice the system uses, since it also records the ef- fect of each utterance on the decision of which bit of information to relay, which topic to explore next. This surprisingly simple information struc- ture can successfully model this important kind of dialogue, as we argue here, and it also makes it rel- atively easy to implement new applications cover- ing other domains in this style—information about companies, for example, or more generally about structured fields of knowledge. The remainder of this paper is organised as fol- lows. Section 2 elaborates on the peculiarities of “information-seeking chats”, and presents our taxonomy of dialogue acts. A discussion of the problems extant approaches to dialogue modelling would have with this kind of dialogue leads over to Section 3, where we describe our approach, and the prototypical implementation. After dis- cussing in Section 4 related attempts to reduce dialogue management to representations of task- knowledge, Section 5 sketches how our dialogue manager fits in with the other modules of the sys- tem that is under development. Section 6 finally discusses evaluation issues, and further work.