A Personality Detection Model Based on Language Style and Contrastive Learning

Launched based on data up to October 2024, personality detection has become a significant area of recent research in psychology, human-computer interaction, and social media analysis. Yet existing computational psychology frameworks ignore the role of writing style in behavioral prediction tasks. Further, a single pre-trained model is still used in many studies to extract semantic information from texts. To overcome these limitations, we propose a novel personality detection approach, styleSem, which jointly leverages writing style and deep semantics. We first apply large language models (LLMs) to extract 11 different writing style features, then incorporate a number of pre-trained models with contrastive learning to enhance semantic aspects. We perform experiments on two common benchmark datasets, Essays and MyPersonality. On the Essays dataset, the new method showed an improvement of 1.74% on average compared to the current state of art algorithm. On MyPersonality dataset, it obtains average accuracy 2.42% higher than top-performing method.