TED: Towards Explainable and Holistically Aligned Empathetic Dialogue Generation

Abstract

While current empathetic dialogue systems can produce fluent responses, they typically lack a transparent reasoning process. This opacity presents a major challenge, as empathetic language models often treat affective and cognitive empathy as disconnected tasks, leading to responses that are emotionally appropriate but logically flawed. Furthermore, their training objectives are misaligned with the complex nature of human empathy, often relying on simplistic text-similarity metrics. To address these issues, we propose a novel method \textbf{t}owards \textbf{e}xplainable and holistically aligned empathetic \textbf{d}ialogue (TED). Our method uses a two-stage training strategy. First, a supervised fine-tuning (SFT) stage teaches the model an explicit chain-of-reasoning. The chain-of-reasoning process compels the model to articulate its emotional and situational analysis before generating an empathetic response, making the inference process transparent and unifying affective and cognitive empathy. Second, we introduce a component-wise relative policy optimization (CRPO) algorithm to refine the SFT-initialized model. The algorithm structures the optimization around multiple semantic objectives by grouping individual reward functions into distinct components, such as logical consistency, emotional alignment, response quality and safety. Critically, policy advantages are calculated independently within each component before being aggregated into a final learning signal. This mechanism encourages the policy to balance trade-offs between empathetic skills and foster holistic development rather than over-optimization on simplistic metrics. Experiments on two multi-modal datasets demonstrate that our method enhances model interpretability while significantly improving response quality. Human judges rated our model’s responses more logically sound, emotionally resonant, and helpful than strong baselines.