CoMA: Condition-Modulated Context–Knowledge Attention for Empathetic Dialogue Generation

Empathetic response generation requires a model to balance emotional appropriateness with contextual grounding. However, the performance of emotion recognition remains relatively limited, which may lead to cascading errors in generating empathetic responses. To avoid error propagation from inaccurate emotion prediction and exploit emotion information more effectively, we propose CoMA (Condition-Modulated Context–Knowledge Attention), an empathetic response generation framework that uses condition from emotion and context representations as modulation signals during decoding. The condition representation is transformed into modulation weights to adjust attention logits, enabling the model to regulate the selectivity of the attention distributions over dialogue context and commonsense knowledge. In addition, CoMA employs a multi-decoder architecture with two separate branches for context and knowledge, whose representations are then fused to generate the final response. Experiments on the EmpatheticDialogues dataset show that CoMA achieves better performance than baseline models in both automatic and human evaluations. On EmpatheticDialogues, CoMA achieves a PPL of 20.47, BLEU of 2.72, Dist-1 of 0.86, and Dist-2 of 4.12. We further conduct an additional evaluation on the ESConv dataset, where CoMA obtains a PPL of 46.80, BLEU of 2.19, Dist-1 of 0.79, and Dist-2 of 3.13. These results indicate that using reliable emotional and contextual information to modulate attention, together with separate context–knowledge decoding, is effective for empathetic response generation across different conversational settings.