Zero-shot Personality Perception From Facial Images

Personality perception is an important process that affects our behaviours towards others, with applications across many domains. Automatic personality perception (APP) tools can help create more natural interactions between humans and machines, and better understand human-human interactions. However, collecting personality assessments is a costly and tedious task. This paper presents a new method for zero-shot facial image personality perception tasks. Harnessing the latent psychometric layer of CLIP (Contrastive Language-Image Pre-training), the proposed PsyCLIP is the first zero-shot personality perception model achieving competitive results, compared to state-of-the-art supervised models. With PsyCLIP, we establish the existence of latent psychometric information in CLIP and demonstrate its use in the domain of personality computing. For evaluation, we compiled a new personality dataset consisting of 41800 facial images of various individuals labelled with their corresponding perceived Myers Briggs Type Indicator (MBTI) types. PsyCLIP achieved statistically significant results (p<0.01) in predicting all four Myers Briggs dimensions without requiring any training dataset.