MODELING THE PRODUCTION OF COVERBAL ICONIC GESTURES BY LEARNING BAYESIAN DECISION NETWORKS

摘要

Expressing spatial information with iconic gestures is abundant in human communication and requires transforming information about a referent into resembling gestural form. This transformation is barely understood and hard to model for expressive virtual agents because it is influenced by the visuospatial features of the referent and the overall discourse context or concomitant speech and its outcome varies considerably across different speakers. We use Bayesian decision networks (BUN) to achieve such a model. Different machine learning techniques are applied to a data corpus of speech and gesture use in a spatial domain to investigate how to learn such networks. Modeling results from an implemented generation system are presented and evaluated against the original corpus data to find out how BDNs can be applied to human gesture formation and which structure learning algorithm performs best.