Combining a data-driven and a theory-based approach to generate culture-dependent behaviours for virtual characters

Abstract

To incorporate culture into intelligent systems, there are two approaches that are commonly proposed. Theory-based approaches that build computational models based on cultural theories to predict culture-dependent behaviours, and data-driven approaches that rely on multimodal recordings of existing cultures. Based on our former work, we present a hybrid approach of integrating culture into a Bayesian Network that aims at predicting culture-dependent non-verbal behaviours for a given conversation. While the model is structured based on cultural theories and theoretical knowledge on their influence on prototypical behaviour, the parameters of the model are learned from a multimodal corpus recorded in the German and Japanese cultures. The model is validated in two ways: With a cross-fold validation we estimate the power of the network by predicting behaviours for parts of the recorded data that were not used to train the network. Secondly we performed a perception study with virtual characters whose behaviour is driven by the calculations of the network and are rated by members of the German and Japanese cultures. With this chapter, we aim at giving guidance for other culture-specific generation approaches by providing a hybrid methodology to build culture-specific computational models as well as potential approaches for their evaluation.