Appraisal Inference from Synthetic Facial Expressions

Introduction

In the field of affective computing there is a strong emphasis on automatic recognition of emotion via facial expression (Gunes, 2010; Valstar, Mehu, Jiang, Pantic, & Scherer, 2012). This mirrors a long-standing interest in psychological work on the recognition of emotion, mostly in the face but also in voice, gestures, or movement. In general, actor portrayals of prototypical emotional expressions are used for recognition studies with lay observer-raters. The analyses are mostly performed by determining the percentage accuracy and/or computing confusion matrices for discrete basic emotions identified with emotion words such as sad, angry, fearful, or joyful (or in some cases, using a dimensional approach in terms of positive or negative valence or low vs. high arousal of the respective expression. In this article, a different approach is proposed, arguing that the emotion recognition process may be quite different from simple matching of faces with labels Rather, it is suggested that it consists of a process of inference based on lower-level information. Concretely, the argument is that observers first infer the expressor's appraisal (i.e., the evaluation of the significance of a given event and the potential for action) followed by the use this information to deduce the likelihood of particular emotions or emotion blends. The study reported here represents a first step in validating this notion, seeking to obtain evidence that judges can reliably infer appraisal from facial action patterns systematically produced via synthesis in avatar faces. The results of this work provide important information on the nature of the psychophysiological processes as well as for the attempt to develop automatic recognition and classification algorithms.