A Parameter-Based Model for Generating Culturally Adaptive Nonverbal Behaviors in Embodied Conversational Agents

The goal of this paper is to integrate culture as a computational term in embodied conversational agents by employing an empirical data-driven approach as well as a theoretical model-driven approach. We propose a parameter-based model that predicts nonverbal expressions appropriate for specific cultures. First, we introduce the Hofstede theory to describe socio-cultural characteristics of each country. Then, based on the previous studies in cultural differences of nonverbal behaviors, we propose expressive parameters to characterize nonverbal behaviors. Finally, by integrating socio-cultural characteristics and nonverbal expressive characteristics, we establish a Bayesian network model that predicts posture expressiveness from a country name, and vice versa.