Abstract The widespread use of social media has facilitated the recognition of personality from user-generated online content. While numerous applications exist across diverse domains, such as recommender systems, most current studies focus on superficial, statistical, and explicit user content, thereby neglecting latent knowledge. In this study, we propose a method for uncovering latent psycholinguistic understanding at deeper levels of user data to enhance personality prediction through natural language processing. The proposed approach leverages fine-tuning of a domain-specific Bidirectional Encoder Representations from Transformers (BERT) model for sentence-level feature extraction and enriches the output by incorporating emotional information. This process emphasizes salient words through a single-way attention mechanism. Our single-way attention mechanism propagates information from highlighted words to the overall extracted knowledge. Subsequently, using the embeddings from the previous stage as node features, we construct a graph. A dynamic, task-oriented learning approach is then employed to determine the graph edges, using a neural network to connect different pairs of nodes. Finally, a graph neural network is combined with a classifier to predict personality traits. Experimental results demonstrate the effectiveness of the proposed model, achieving 80.27% accuracy on the Essays dataset and outperforming existing approaches. Furthermore, several ablation studies were conducted to investigate the impact of various components and parameters of the proposed architecture.
Bajestani et al. (Mon,) studied this question.