Abstract:
As computer systems become more sophisticated, they can assist human beings in performing more complicated tasks. One such application is in the development of advice-giving systems in which the computer system plays the role of a consultant and helps a user develop a plan to achieve his goal. However, since the computer system and its user each have private beliefs about the world and about each other, it is almost inevitable that disagreements will arise between them during their interaction. In order for the computer system to appear cooperative, it must be able to detect these disagreements, determine whether or not they are obstacles to accomplishing the task at hand, and if so, attempt to resolve these disagreements in order to successfully complete its task. In this talk, I will describe a recursive "Propose-Evaluate-Modify" framework for modeling collaborative dialogues. I will then elaborate on the "evaluate" and "modify" phases of this process and discuss: 1) the aspects of a user proposal that a collaborative system must take into account when evaluating the proposal, and 2) the key issues that must be addressed when such a system attempts to generate cooperative responses to resolve a relevant conflict detected in the user proposal. Finally, I will show some sample responses generated by my implemented system CORE, and a summary of the cooperativeness and coherence of CORE's responses judged by human subjects.