知识对话是指对话系统利用外部知识信息，使聊天内容更加丰富、准确，这对提升用户体验是非常重要的，近年来受到学术界和工业界的广泛关注。
Knowledge grounded dialogue means that the dialogue system uses external knowledge information to make the conversation more engaging and factually correct. This is very important for improving user engagement and has gained a lot of attention from both academia and industry in recent years.
为了解决静态知识的丰富性、时效性和个性化问题，我们提出了一个全新的知识对话任务——搜索信息增强的对话(SINC)。对话系统在对话的过程中动态地搜索外部知识信息，并将搜索知识用于回复生成中。为此我们建设了外部知识搜索API，可以根据给定query和用户地理位置实时搜索各类通用知识、动态知识和个性化知识，同时我们利用这个API人工建设了用于该任务研究的对话数据集DuSinc。
To address the lack of richness, timeliness, and personalization of static knowledge, we propose a novel knowledge grounded dialogue task called Search INformation augmented Conversation (SINC). The dialogue system dynamically searches for external knowledge information in the process of conversation and uses the searched knowledge in response generation. To this end, we have built an external knowledge search API, which can search various general knowledge, dynamic knowledge, and personalized knowledge in real-time according to the given query and user geographic location. At the same time, we use this API to manually build a dialogue dataset named DuSinc for this task research.
本次竞赛中，我们主要从以下两个子任务评测系统的知识对话能力：1）Query生成任务：给定多轮对话历史，生成用于查询搜索引擎的Query；2）回复生成任务：给定文本知识与多轮对话历史，生成合适的对话回复。
In this challenge, we mainly evaluate the knowledge grounded dialogue ability of the system from the following two subtasks: 1) Query Generation Task: given dialogue history, generate a query for querying the search engine; 2) Response Generation Task: given text knowledge and dialogue history to generate appropriate dialogue responses. |
