Takes Two to Know One: An Approach for Personality Prediction from Dyadic Conversation Based on Text

Abstract

Personality strongly influences how individuals communicate, interact, and collaborate with others. However, accurately predicting personality from conversation remains challenging due to limited interaction time, incomplete context, and the dynamic nature of communication. Prior text-based computational approaches have largely focused on monologic text and self-rated personality with comparatively little attention to perceived personality inferred from dyadic interactions, which offer richer —though more complex— signals from the interplay between conversational partners. In this study, we investigated how Big Five personality traits can be inferred from dyadic dialogue using a dataset of 24 participants engaged in a collaborative survival task. Three key research questions were addressed: (i) how much dialogue was required to reliably predict perceived personality, (ii) whether incorporating a partner’s dialogue improved prediction accuracy, and (iii) which linguistic features were most effective in modelling personality. We proposed a weighted fusion scheme that incorporated a conversational partner’s linguistic features and compared the performance against linguistic and temporal features, as well as BERT embeddings, derived from an individual alone, to explore the role of interaction in personality prediction. The results from leave-one-participant-out binary classification showed that personality can be predicted with 75-80% accuracy within as few as 25–35 dialogue turns. Dyadic fusion improved the accuracy by up to 35-50% over individual features for openness, neuroticism and conscientiousness, and by up to 50% over BERT embeddings across all five traits. The most effective features also aligned with psychological theories. Our findings highlight the importance of interdependence between conversational partners in personality prediction.