A Robust System for Natural Spoken Dialogue

A Robust System for Natural Spoken Dialogue James F. Allen, Bradford W. Miller, Eric K. Ringger, Teresa Sikorski Dept. of Computer Science University of Rochester Rochester, NY 14627 {james, miller, ringger, sikorski)@cs.rochester.edu http : //www. cs. rochester, edu/research/trains / Abstract This paper describes a system that leads us to believe in the feasibility of constructing natural spoken dialogue systems in task-oriented domains. It specifically addresses the issue of robust interpre- tation of speech in the presence of recognition errors. Robustness is achieved by a combination of statistical error post-correction, syntactically- and semantically-driven robust parsing, and extensive use of the dialogue context. We present an evaluation of the system using time-to-completion and the quality of the final solution that suggests that most native speakers of English can use the system successfully with virtually no training. 1. Introduction While there has been much research on natural dialogue, there have been few working systems because of the difficulties in obtaining robust behavior. Given over twenty years of research in this area, if we can't construct a robust system even in a simple domain then that bodes ill for progress in the field. In particular, without some working systems, we are very limited in how we can evaluate the worth of different models. The prime goal of the work reported here was to demonstrate that it is feasible to construct robust spoken natural dialogue systems. We were not seeking to develop new theories, but rather to develop tech- niques to enable existing theories to be applied in practice. We chose a domain and task that was as simple as possible yet couldn't be solved without the collaboration of the human and system. In addition, there were three fundamental requirements: • the system must run in near real-time; • the user should need minimal training and not be constrained in what can be said; and • the dialogue should have a concrete result that can be independently evaluated. The second constraint means we must handle natural dialogue, namely dialogue as people use it rather than a constrained form of interaction determined by the system (which is often called a dialogue). We can only control the complexity of the dialogue by controlling the complexity of the task. Increasing the task complexity naturally increases the complexity of the dialogue. This paper reports on the first stage of this process, working with a highly simplified domain. At the start of this experiment in November 1994, we had no idea whether it was possible. While researchers were reporting good accuracy (upwards of 95%) for speech systems in simple question-answering tasks, our domain was considerably different with a much more spontaneous form of interaction. We also knew that it would not be possible to directly use general models of plan recognition to aid in speech act interpretation (as in Alien & Perrault, 1980, Litman & Allen 1987, Carberry 1990), as these models would not lend themselves to real-time processing. Similarly, it would not be feasible to use general planning models …

Content truncated to 500 words