真实的推荐系统通常基于与用户的交互，以提升推荐精度和服务质量。对话式推荐系统基于自然语言，通过多轮对话来深入理解用户，以实现更精准的用户偏好建模与候选物品推荐，并通过生成高质量的回复来提升用户体验。在真实的应用场景中，往往涉及到用户在多个领域中的推荐。然而，现有的对话式推荐系统数据集通常只涉及单个领域物品的推荐，无法满足跨域推荐任务的需求。基于此，本文提出了一个跨领域的对话式推荐系统数据集CD-ReDial。我们利用预训练语言模型与用户评论、知识图谱等外部知识，经过从对话流到对话两阶段的数据构建过程，并采用基于主动学习的半自动标注方式，以较低成本获取高质量的数据集。实验结果表明，我们构建的数据集能够有效地解决跨域推荐的问题。

Real-world recommender systems often rely on interactions with users to improve recommendation accuracy and service quality. Conversational recommender systems enable in-depth understanding of users for more accurate user preference modeling and candidate item recommendation based on natural language and improve user experience by generating high-quality responses. In real-world application scenarios, recommendations are often needed in multiple  domains. However, existing conversational recommender system datasets usually only involve single-domain item recommendation and cannot meet the requirements of cross-domain recommendation tasks. In this paper, we propose a cross-domain conversational recommender system dataset CD-ReDial. We leverage pre-trained language models and external knowledge such as user comments and knowledge graph to construct the dataset through a two-stage process: dialog flows à dialogs, and adopt a semi-automatic annotation based on active learning to acquire high-quality data at a low cost. Experimental results demonstrate the effectiveness of our dataset in addressing the problem of cross-domain recommendation.