Three-way system-user-expert interactions help you expand the capabilities of an existing spoken dialogue system

摘要

Data collection and use is absolutely necessary to contemporary spoken dialogue system research. Yet much of it is done either at the beginning of a project with human-human conversation for a particular task, or done with a "frozen" system as an evaluation technique. As system development proceeds, system-user dialogue gets further and further apart from the original human-human dialogues due to the inevitable (and often desired) adaptation of the task to the strengths and limitations of the computer conversational partner. At the time of any particular evaluation, only the features present in the dialogue system will be evaluated. This may reveal which areas need further work, but sheds precious little light on how to expand the system to carry out the necessary novel subdialogues. This paper describes a new methodology to expand the capabilities of an existing dialogue system. The central technique uses three-way dialogues between a system, user, and expert as successive approximations of what people do. We describe three substantive uses of the methodology, in various systems: an intelligent tutoring system for children's oral reading; a spoken interface to a robotic helicopter simulation; and, an intelligent procedure assistant for astronauts. In each case, three-way dialogues allowed us to determine how the dialogue system should behave in the new parts of the dialogue under study, and then to add appropriate new functionality to the dialogue system itself. The three-way dialogue methodology is a powerful new technique for spoken dialogue system development.