【摘要】 情感支持对话任务旨在通过对话的方法帮助用户进行情绪疏导，克服其所面临的问题。现有工作仅基于对话历史选择支持策略，用户情绪反馈对策略选择的影响未得到充分利用。因此，该文提出一种基于用户预期情绪感知策略选择的情感支持对话模型(Lookahead Emotion-aware Strategy Selection Model for Emotional Support Conversation, EmoESC),通过模拟用户对不同支持策略产生的反应得到预期情绪，利用强化学习的方法将策略使用前后用户的情绪变化作为奖励，选择使用户预期情绪更佳的支持策略。同时，为了增加回复与用户问题之间的相关性，利用用户问题描述增强回复生成，从而有效缓解用户负面情绪并提供帮助。在ESConv公开数据集上进行的实验表明，EmoESC模型在5项自动评估指标和3项人工评估指标上取得了比其他基准模型更好的效果。更多还原

【Abstract】 Emotional support conversation aims to help users alleviate negative emotions. Existing methods usually select support strategies only based on dialogue history without capturing the influence of users’ emotion feedback. To address this issue, this paper proposes a Lookahead Emotion-aware Strategy Selection Model for Emotional Support Conversation(EmoESC). Firstly, a user simulator is employed to simulate the user’s responses toward different support strategies. Then, a strategy selector chooses the strategy that best improves user’s negative emotion by reinforcement learning. Finally, the user’s situation information is used to enhance the response generation to effectively soothe the user’s negative emotion and provide support. The experimental results on the ESConv dataset show that the EmoESC model achieves better results than other baseline models.更多还原