From Static Traits to Dynamic Affect: A Reasoning-and-Rewriting Framework for Personality Dialogue

ACL ARR 2026 January Submission5659 Authors

Keywords: Personality-aware Dialogue;Dialogue Systems;Emotional Dialogue Generation;Controllable Text Generation;Big Five Personality Traits

Abstract: While large language models (LLMs) have achieved remarkable fluency, they often struggle to maintain stable personas and exhibit an inherent “neutrality bias” resulting in emotionally flat and insufficiently personalized interactions. To address these challenges, we propose a novel personality-driven dialogue framework that explicitly models the reasoning chain from Big Five traits to affective realization. Specifically, we decouple the generation process into three distinct stages: Personality-Aware Emotion Prediction, which induces target affective states from structured trait profiles and conversational context; Affective Refinement, where an initial response is rewritten via Direct Preference Optimization (DPO) to amplify emotional intensity; and External Verification, which closes the loop through an independent emotion recognition module. Our framework effectively compensates for the affective limitations of base LLMs while preserving their semantic coherence. Experimental results on personality-annotated corpora demonstrate that our approach significantly outperforms strong baselines in both emotional fidelity and personality consistency, as validated by automatic metrics and human evaluation.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Conversational Modeling;Evaluation and Metrics;Task-Oriented

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 5659