Emotion Change Reasoning in Chinese Multi-Turn Dialogue via Multi-Task Parameter-Efficient Fine-Tuning of Large Language Models

Emotion change reasoning in multi-turn dialogues represents a critical challenge for developing empathetic human-computer interaction systems, requiring a deep understanding of the causal mechanisms behind emotional transitions. This paper addresses the complexity of this task in Chinese conversational contexts through a novel Multi-Task Quantized Low-Rank Adaptation framework based on large language models. The proposed approach integrates three fundamental subtasks of textual stimulus identification, cognitive appraisal generation, and emotional response prediction into a unified modeling architecture that enables joint learning and holistic reasoning along the stimulus-cognition-emotion causal chain. Building upon the Qwen3-32B foundation model, parameter-efficient fine-tuning that combines 4-bit quantization with low-rank adaptation significantly reduces computational requirements while maintaining strong performance. Experimental results on the Chinese emotional dialogue dataset demonstrate superior performance across multiple evaluation metrics, achieving a METEOR score of 0.5024 and a BERTScore of 0.8100, substantially outperforming existing baseline models and securing first place in the 5th China Conference on Affective Computing evaluation. These findings validate the effectiveness of the framework in capturing the causal dynamics of emotional change while providing a practical and efficient paradigm for emotion reasoning in Chinese dialogue systems with real-world applicability. The implementation of the MT-QLoRA framework is publicly available at https://github.com/lidayuls/EmotionChangeReasoning .