基于强化学习的情感对话回复生成

Emotional conversation generation based on reinforcement learning

近年来,研究者们致力于提升对话系统的情感智能,但都忽略了对话中的反馈因素,模型容易生成无趣简单的通用回复。针对此问题,提出一种基于强化学习的情感对话生成方法EC-RL(emotional conversation reinforcement learning),根据对话上文生成指定情感的回复语句,从内容质量和情感两个方面评估对话的未来奖励,增强生成语句的连贯性,采用NLPCC 2017情感对话数据集,实验证明,与现有方法相比,所提模型可以实现情感可控的对话回复生成,文本内容通顺流畅且多样性高。

In recent years,researchers have been dedicated to enhancing the emotional intelligence of dialogue systems,yet have overlooked the feedback elements within dialogues.These models tend to produce uninteresting and simplistic general re‑sponses.To address this issue,we propose an emotional conversation reinforcement learning(EC-RL)approach.This method gen‑erates response statements with specified emotions based on the context of the conversation,and evaluates the future reward of the dialogue from both content quality and emotional aspects to improve the coherence of the generated statements.Utilizing the NLPCC 2017 emotional dialogue dataset,experiments demonstrate that,compared to existing methods,our proposed model enables emotionally controllable dialogue response generation,producing fluent and diverse text content.