Reinforcement Learning-Driven Adaptive Emotion-Cause Analysis in Conversation

The research on Emotion-Cause Analysis in Conversation (ECAC) seeks to reveal the emotions expressed in utterances and their corresponding causes. Currently, ECAC-related studies typically rely on prompt engineering and fine-tuning to enhance the models’ adaptability to tasks. However, these studies often employ static prompts or fixed fine-tuning templates to handle all samples, which often leads to limited generalization and interference from irrelevant information when facing complex and diverse conversational scenarios. To address them, this paper proposes a Reinforcement Learning-Driven Adaptive Emotion-Cause Analysis in Conversation (RLDA) framework. RLDA first converts the pre-trained model fine-tuning problem into a policy function optimization problem based on reinforcement learning, reducing the dependence on the quality of training data and enhancing the model’s adaptability across different conversations. Second, it reduces the interference from irrelevant prompts by selecting few-shot prompts that are tailored to the target conversation. Finally, it utilizes a majority voting mechanism to stabilize the output of the Large Language Model (LLM). Experimental validation shows that RLDA performs competitively on the RECCON-DD dataset and demonstrates the generalization capabilities in long, emotionally complex conversations on the RECCON-IE dataset.