【摘要】 共情回复生成旨在理解对话中用户的经历与感受并表达出合理的回复.心理学理论认为，角色是人格的外在表现，与共情密切相关.然而，现有工作主要关注共情的认知因素和情绪因素，忽略有益于共情的角色因素，导致缺少个性化的共情回复.为了解决该问题，文中提出角色增强的共情回复生成模型(Persona-Enhanced Empathetic Response Generation Model, PERG).首先，为了有效利用角色信息，提出角色增强编码模块，通过编码器捕获上下文、情境及角色信息的深层语义关系，结合上下文和情境筛选角色信息，提升模型对说话者与回应者角色的理解，增强共情能力.然后，在角色调控解码模块中，设计基于多解码器融合的调控机制，有效结合角色信息，调节上下文和情境对共情回复的影响，生成高度个性化的共情回复.在公开的共情回复EmpatheticDialogues数据集上的实验表明，PERG在多个指标上均取得较优值.更多还原

【Abstract】 Empathetic response generation aims to understand the experiences and feelings of users in conversations and provide appropriate responses. Psychological theories suggest that roles serve as an external manifestation of personality and are closely related to empathy. However, existing research primarily focuses on the cognitive and emotional factors of empathy while neglecting role factors that are beneficial to empathy, resulting in a lack of personalized empathetic responses. To address this issue, a persona-enhanced empathetic response generation model(PERG) is proposed. A persona-enhanced encoding module is introduced to capture deep semantic relationships among context, situation and role information through an encoder. By filtering role information based on context and situation, the understanding of the speaker′s and responder′s roles by the model is significantly improved, and thereby enhancing its empathetic capabilities. In the persona control decoding module, a multi-decoder control fusion mechanism is designed. The role information is effectively combined to regulate the impact of context and situation on empathy responses, generating highly personalized empathetic responses. Experiments on EmpatheticDialogues dataset indicate that PERG achieves superior performance in multiple metrics.更多还原