A Framework for Automatic Personality Recognition in Dyadic Interactions

Research has shown that the way in which an individual interacts with others contains vital cues for recognising their real personality traits. The ability to recognise and adapt to the personality of users is key to developing more intelligent social robots, especially in real-world scenarios. However, most methods for personality recognition focus on apparent personality recognition of individuals in isolated settings. In this work, we propose the first multi-modal framework for human behaviour primitives-based automatic real personality recognition in dyadic interactions. It leverages the use of the spectral representations of behavioural primitives to exploit the temporal nature of the data whilst retaining as much vital information pertaining to personality as possible. We experiment on a range of standard fusion methods to evaluate their effectiveness at combining information from multiple modalities and both interactants in a dyadic interaction. At the multi-subject level, our attention-based fusion approach using a multimodal transformer enabled with cross-subject attention was the most successful. The experimental results show that our approach improved on the previous state-of-the-art on the UDIVA dataset by up to 46%.