Auto-generating Virtual Human Behavior by Understanding User Contexts

Abstract

Virtual humans are most natural and effective when it can act out and animate verbal/gestural actions. One popular method to realize this is to infer the actions from predefined phrases. This research aims to provide a more flexible method to activate various behaviors straight from natural conversations. Our approach uses BERT as the backbone for natural language understanding and, on top of it, a jointly learned sentence classifier (SC) and entity classifier (EC). The SC classifies the input into conversation or action, and EC extracts the entities for the action. The pilot study has shown promising results with high perceived naturalness and positive experiences.