DuRecDial 2.0中英双语平行对话推荐数据集

DuRecDial 2.0是第一个大规模中英双语平行的对话推荐数据集，包含5种对话类型(闲聊、对话推荐、知识对话、任务对话、QA)、6个领域(明星、电影、音乐、美食、POI、天气)、16.5k个对话和255k个utterance，采用Wizard-of-Oz方式人工构建。在每个对话中，推荐者(bot)使用丰富的交互行为，主动引导一个多类型对话不断接近推荐目标。DuRecDial 2.0旨在考察模型是否可以在对话过程中基于用户兴趣以及用户的实时反馈，主动给用户做出合理的推荐。DuRecDial 2.0可支持单语言对话推荐、多语言对话推荐和跨语言对话推荐任务。DuRecDial 2.0 is the first large-scale Chinese-English parallel conversational recommendation dataset. It encompasses 5 dialogue types (chit-chat, conversational recommendation, knowledge-grounded dialogue, task-oriented dialogue, QA), 6 domains (celebrities, movies, music, food, POI, weather), 16.5k dialogues and 255k utterances, and is manually constructed using the Wizard-of-Oz paradigm. In each dialogue, the recommender bot leverages a variety of interactive behaviors to actively steer a multi-type conversation toward the recommendation target. DuRecDial 2.0 aims to evaluate whether a model can proactively generate reasonable recommendations for users during the conversation based on their interests and real-time feedback. It supports monolingual, multilingual and cross-lingual conversational recommendation tasks.