A computational model of turn-taking in discourse

Abstract

This thesis presents a computational model for turn-taking applicable to dialog-based interactive systems. A general, three-part model of the internal architecture of a turn-taking agent is presented, accounting for why an agent should take a turn, what it should contribute, and when it should act. For deciding what to say, the next-action constraint optimization problem is developed, which allows an agent to order multiple goals using an interruptible local search algorithm sensitive to the time-constraints imposed by dynamic dialog environments. For deciding when to speak, a formal account of turn-taking is given, based on the situation calculus, and accompanied by a state-based algorithm that tells an agent when it should take a turn based in part on the status of its processing. Two prototype implementations of course-advising systems are presented, to illustrate how interactive systems can be analyzed using a checklist for system designers derived from the turn-taking model. In developing a computational model for turn-taking, this thesis shows how to bring models for written discourse one step closer to spoken dialog, and thus advances the state of the art in natural language discourse theory and in natural language generation.