Modelling Goal Oriented Referential Dialogue using Discourse Coherence, Cognitive Modelling, and RL
Wednesday, October 07, 2020, 12:30pm - 02:00pm
Speaker: Baber Khalid
Location : Remote via Webex
Dr. Matthew Stone (advisor), Dr. Yongfeng Zhang, Dr. Abdesalam Boularias, Dr. Zheng Zhang
Event Type: Qualifying Exam
Abstract: Most of the dialogue literature either focuses on evaluation of utterances on their own creating problems in reference resolution or build end-to-end dialogue systems which are hard to customize. We present a dialogue system which represents discourse state as a graph of coherence relations. We use a referential dialogue task in which a director has to describe an ambiguous target to a matcher to show that a coherence-based representation can be successfully used to model dialogue contributions at discourse level and accomplish reference resolution of ambiguous referents. We use data-driven methods to implement individual dialogue modules, considerably reducing human effort. This representation is coupled with a RL method to learn flexible dialogue strategies. Secondly, we show that by combining this method with cognitive modelling we can learn context-sensitive clarification strategies which improve interactive experience with human subjects and better task success rate.