Integrating Long-term and Short-term Perspectives for Textual Personality Detection

Personality detection infers individuals’ personality characteristics from their textual expressions, which serves as a foundation for constructing and guiding personas in large language models (LLMs). Existing studies typically focus on either the long-term or the short-term perspective, neglecting their joint influence in personality modeling. Motivated by psychological evidence that personality is both stable and adaptive, we propose a Dual Enhanced Network (DEN) that jointly models long-term and short-term personality representations. The long-term module captures consistent linguistic and behavioral patterns reflecting enduring traits, while the short-term module encodes contextual variations across individual posts to capture transient states. A bi-directional interaction mechanism further integrates these two perspectives into a unified representation. Extensive experiments demonstrate that DEN outperforms strong baselines and highlight the importance of jointly modeling long-term and short-term cues for comprehensive personality modeling.