PhD 2023F: Sarah Finch
Enriching Dialogue Agents with Enhanced Experience Understanding through Social Commonsense Integration
Sarah Finch
Date: 2023-10-13 / 3:30 ~ 4:30 PM
Location: MSC E306
Abstract
Human social interaction often revolves around the exchange of personal experiences. To achieve more natural human-bot dialogues, it is imperative to emulate this thematic focus. This necessitates chat-oriented dialogue agents with a profound comprehpropension of human experiences. Social commonsense is an integral component towards better understanding of human experience, since it captures the building blocks of human experience including motivations, desires, causes, effects, etc. I will discuss how experience-centered conversations influenced by commonsense have found success through rule-based dialogue agents, albeit on a limited scale. I will also introduce an evaluation framework that exposes that the prevalent deficiencies in neural dialogue models are related to the understanding of human experiences, leading to commonsense lapses, partner contradictions, and self-inconsistencies. Thus, there exists a pressing need for improved social commonsense understanding to facilitate more intelligent experience-oriented conversations in neural dialogue agents. This thesis proposal outlines two key contributions towards addressing this need: (1) the development of an enhanced social commonsense model tailored to dialogue contexts and (2) the integration of this commonsense as contextual guidance for response generation.