A Computational Model of Dialogue Initiative in Collaborative Discourse

This paper presents a model of dialogue initiative within a collaborative discourse. The importance of varying initiative between participants is validated both in theory and in experiments. For example, it was found that one initiative setting algorithm was up to 50% better than if initiative is random. Furthermore, a method for varying levels of initiative between participants is presented which relies on a relatively simple user model. 1. Motivation Two recently implemented voice dialogue systems utilize the principle that discourse is a by-product of cooperative problemsolving[1, 2]. The computer participant in both of these systems proceeds by attempting to prove or solve some goal, and when the system is blocked, it has the option of using dialogue to obtain the necessary information in order to continue its proof. This dialogue mechanism has been described as the missing axiom theory of discourse[3]. However, efficient collaboration requires efficient management of dialogue control. Each of the cooperative problem-solvers must take control or release it at appropriate times to maximize the joint efficiency of their collaborative problem-solving. This paper will explore various mechanisms by which a problem-solver may set and possible reset its initiative levels during collaborative problem-solving between two participants. 2. The Collaborative Model The model developed assumes two beings are working collaboratively to solve a problem each using its own private knowledge and user model of the other participant as illustrated in Figure 1. A set of facts about the problem is distributed in some manner between the participants, and the discourse involves the interactions needed to prove the subgoals required to solve the global problem. The process involves each participant attempting to solve the problem alone and making occasional requests of the other when particular subgoals fail. When a participant receives a request from the other, it has the option of trying to solve the requested goal and passing it back, or ignoring the request and pursuing its own goal. Plan ? Plan Dialogue