Dept. of Computer Science and Electrical Engineering
Oregon Health & Science University
We envision that next-generation spoken dialogue systems will be supporting a complex user goal and multiple parallel tasks, which requires the system and the user to jointly direct the flow of conversation. The problem, however, is that there lacks an effective model of directing the flow of conversation. This thesis research aims to develop such a model for next-generation spoken dialogue systems. We started with conventions actually used in human-human dialogue, which are natural for users to follow and probably also efficient in problem-solving. An annotation framework, DialogueView, was established to allow for the investigation of complex interaction in dialogue. A series of empirical studies on two corpora, the Trains and the MTD, were then conducted to understand people's initiative behavior of directing the conversation flow. We first examined people's initiative behavior in decomposing a complex goal into sub-goals and achieving each of them in the Trains domain. We found that initiative is subservient to discourse goal. We next examined people's initiative behavior in switching the conversation to a more urgent task. We found that conversants strive to switch tasks at a less disruptive place; but where they cannot, they exert additional effort to signal the task switching, such as increasing pitch. We finally examined people's behavior on initiative conflicts, where both conversants try to direct the conversation at the same time. We found that conversants try to avoid initiative conflicts; but when initiative conflicts occur, they are efficiently resolved with simple linguistic devices such as volume. Computer simulation experiments were also conducted to better understand the underlying benefits of using the human conventions. Our findings on human-human dialogues have important implication for building next-generation spoken dialogue systems by (1) guiding the system when to show initiative, and when to let the user show initiative; (2) guiding the system when and how to switch to a more urgent task, and to understand the user's switch; and (3) guiding the system how to resolve and repair initiative conflicts.
OGI School of Science and Engineering
Yang, Fan, "Directing the flow of conversation in task-oriented dialogue" (2008). Scholar Archive. 307.