An Application of Reinforcement Learning to Dialogue Strategy Selection in a Spoken Dialogue System for Email

Walker, Marilyn
An Application of Reinforcement Learning to Dialogue Strategy Selection in a Spoken Dialogue System for Email
Journal of Artificial Intelligence Research, Vol 12., pp. 387-416, 2000. (Postscript - 340K )

Abstract: In the past several years, it has become possible to build spoken dialogue systems that can communicate with humans over the telephone in real time. Systems exist for tasks such as finding a good restaurant nearby, reading your email, perusing the classified advertisements about cars for sale, or making travel arrangements. These systems are some of the few realized examples of real time, goal-oriented interactions between humans and computers. Yet in spite of 30 years of research on algorithms for dialogue management in task-oriented dialogue systems, the design of the dialogue manager in real-time, implemented systems is still more of an art than a science. This article describes a novel method using reinforcement learning, and experiments that validate the method, by which a spoken dialogue system can learn from its experience with human users to optimize its choice of dialogue strategy.