Learning Human Emotion Patterns for Modeling Virtual Humans

Abstract

Emotion modeling is a crucial part in modeling virtual humans. Although various emotion models have been proposed, most of them focus on designing specific appraisal rules. As there is no unified framework for emotional appraisal, the appraisal variables have to be defined beforehand and evaluated in a subjective way. In this paper, we propose an emotion model based on machine learning methods by taking the following position: an emotion model should mirror actual human emotion in the real world and connect tightly with human inner states, such as drives, motivations and personalities. Specifically, a self-organizing neural model called Emotional Appraisal Network (EAN) is used to learn from human being's emotion patterns, involving context, events, personality and emotion. Our experiments in a virtual world domain have shown that comparing with other emotion models, EAN has a much higher accuracy in emulating human emotion behaviour by learning from real human data.