Sequential Models for Automatic Personality Recognition from Multimodal Information in Social Interactions

The task of automatic personality recognition has become very popular in recent years and it is considered a difficult one as we are trying to model human behavior that may not be visually obvious. Although state-of-the-art approaches have used deep learning architectures such as Transformers and some techniques such as Neural Architecture Search (NAS), some of these methods disregard valuable temporal information. In this paper, we approach the task by modeling it as a sequential problem, using a bimodal recurrent neural network, and exploiting the visual and textual modalities jointly. We report experimental results obtained in a novel corpus of dyadic interactions, outperforming state-of-the-art for the Extraversion personality trait. Another contribution of this paper is that we also analyze the regression to the mean problem that we think most state-of-the-art approaches could be facing when approaching the personality recognition task.