Context Modulates Internal Representations of Facial Emotion: Evidence from a Genetic Algorithm Approach

Abstract

Affect recognition and communication are critical for everyday social interaction. Traditional emotion research has often assumed that facial expressions reliably reflect internal emotional states and conform to a set of universal expressions. However, emerging evidence suggests that facial expressions are highly individualized and influenced by context. Recent work using genetic algorithms (GAs) has revealed profound individual differences in internal representations of emotion, but these studies have not examined how contextual information modulates such representations. To investigate the joint effects of context and individual differences, we used GAs to enable 12 participants to generate personalized facial expressions matching their subjective interpretation of affect within written contextual scenarios. Standardized scenarios drawn from Howard Schatzs Actors Acting were independently rated across 13 primary emotions. In the MAX condition, each primary emotion was paired with a scenario maximizing a secondary emotion; in the MIN condition, the secondary emotion was minimized; in the WORD condition, participants selected faces based on emotion words alone. Faces were generated within the 199-dimensional shape coefficient space of the Basel Face Database across 6 generations per trial. We found a significant main effect of condition on selection rate (p < .001), with participants selecting more faces under contextual conditions than under WORD. Critically, context differentially modulated specific emotions: Embarrassment, Surprise, and Disgust showed large context effects (all p < .001), whereas Sadness, Awe, and Anger did not. Cosine distance analysis confirmed that face clusters were emotion-specific (d = -7.88) and that their separation varied with context. The GA showed robust convergence across generations (d = 0.61). Strong individual differences were observed (2 = .229), with low inter-participant agreement (mean r = .06). These results support constructionist models of emotion by demonstrating that facial expression representations are context-dependent, with context selectively modulating certain emotions more than others.