An implementation of face-to-face grounding in an embodied conversational agent

Abstract

When people have a face-to-face conversation, they don't just spout information blindly-they work to make sure that both participants understand what has been said. This process of ensuring that what has been said is added to the common ground of the conversation is called grounding. This thesis explores recent research into the verbal and nonverbal means for grounding, and presents an implementation of a face-to-face grounding system in an embodied conversational agent that is based on a model of grounding extracted from the research. This is the first such agent that supports nonverbal grounding, and so this thesis represents both a proof of concept and a guide for future work in this area, showing that it is possible to build a dialogue system that implements face-to-face grounding between a human and an agent based on an empirically-derived model. Additionally, this thesis describes a vision system, based on a stereo-camera head-pose tracker and using a recently proposed method for head-nod detection, that can robustly and accurately identify head nods and gaze state.