Emotions in Socio-cultural Interactive AI Agents

Abstract

With the advancement of AI and Robotics, computer systems have been put to many practical uses in a variety of domains like healthcare, retail, households, and more. As AI agents become a part of our day-to-day life, successful human-machine interaction becomes an essential part of the experience. Understanding the nuances of human social interaction remains a challenging area of research, but there is growing consensus that emotional identity, or what social face a person presents in a given context, is a critical aspect. Therefore, understanding the identities displayed by humans, and the identity of agents and the social context, is a crucial skill for a socially interactive agent. In this paper, we provide an overview of a sociological theory of interaction called Affect Control Theory (ACT), and its recent extension, BayesACT. We discuss how this theory can track fine grained dynamics of an interaction, and explore how the associated computational model of emotion can be used by socially interactive agents. ACT considers the cultural sentiments (emotional feelings) about concepts for the context, the identities at play, and the emotions felt, and aims towards a successful interaction with the aim of maximizing emotional coherence. We argue that an AI agent's understanding of itself, and of the culture and context it is in, can change human perception of an agent from something that is machine-like, to something that can establish and maintain a meaningful emotional connection.