Modeling and Evaluating a Bayesian Network of Culture-specific Non-verbal Behaviors for Virtual Character Dialogues

Human behavior is, amongst others, influenced by cultural background. Thus research on virtual characters that should imitate human behavior, considers cultural background to influence the behavior of agents as well. This paper presents a hybrid approach of modeling culture-specific non-verbal behaviors for virtual character dialogues based on both: theoretical knowledge and empirical data. A Bayesian network model has been established that augments the causal relations between culture, dialog behavior, and non-verbal behavior with empirical data. Therefore, the structure and variables of the network are designed based on models and theories from the social sciences, while its parameters are learned from a video corpus of German and Japanese conversations. A 10-fold-cross-validation has been conducted to validate the model. Having such a model at hand allows further investigation of behavioral dependencies as well as the automatic generation of non-verbal behaviors for virtual character dialogs that reflect prototypical differences based on culture.