知识图谱辅助的垂直领域自动人机对话系统框架研究

Automatic human-computer dialogue technology is one of the key research topics in Artificial Intelligent. However, the existing dialogue systems tend to put focus in the relevance between the user-issued queries and produced replies, but often fail to guarantee the richness of information provided in the systems’ replies. Especially in the scenarios of vertical applications, where people are expecting to receive expert knowledge from the dialog systems. Therefore, there is a pressing need of a complete framework to guide the construction of automatic human-computer dialogue systems in the vertical domain scenarios, which could produce proper replies with both high semantic relevance and sufficient knowledge and information...In this project, we propose a complete framework to guide the construction of automatic human-computer dialogue systems for any specific vertical domain. First, by integrating heterogeneous data from different sources, such as general knowledge graphs and text documents, we construct high-precision and fine-grained domain knowledge graphs as the basic resources. Then, we model the dialogue by taking word representations, the user-issued query, and the context into consideration, aiming at understanding the user’s request. Next, we will automatically determine whether the mechanism of domain knowledge graphs should be triggered or not. If it should be triggered, we would extract the associated relevant entity information from the pre-constructed knowledge graphs. Finally, based on the semantic analyses above, we construct a neural network model to generate controlled dialogues, which could guarantee the existence of associated entity information would appear in the generated reply. To achieve a clearer and more vivid explanation, we will implement a dialogue prototype system for the domain of healthcare information.

自动人机对话技术是人工智能热点问题，但目前该技术只能够满足对话系统的回复与用户当前语句在一定程度上的相关性，却无法保证回复中包含足够的有效信息。尤其是在垂直应用的场景下，人们更希望人机对话系统可以提供有用的领域知识。因此，迫切需要一套完整的框架体系，能够指导构建垂直领域自动人机对话系统，使得系统给出的回复既满足与上下文之间的相关关系，又包含足够的知识与信息。本项目提出了一套完整的框架，旨在指导特定垂直领域自动人机对话系统的构建。此框架首先利用通用知识图谱、文本文档等多源异构数据，整合出高精度细粒度的领域知识图谱，作为基础资源来提供知识与信息；然后，从字表征、用户表述、上下文信息三个方面对当前语境建模，理解用户表述语义，并据此进行领域知识图谱触发判定以及实体联想。本框架将建立神经网络模型实现受控对话生成，保证联想到的实体信息能够出现在回复中。本项目将建立医疗健康领域的人机对话原型系统。

为了使人机对话系统在各类垂直应用场景中可以为用户提供有用的领域知识，本项目提出了一套完整的框架，旨在指导构建垂直领域自动人机对话系统，使得系统给出的回复既满足与上下文之间的相关关系，又包含足够的知识与信息。围绕此目标，本项目从构建领域知识图谱、语义理解和表示学习、个性化问诊系统这三个方面出发，提供了在知识图谱辅助的垂直领域自动人机对话系统构建过程中所需要用到的技术。为了能够将医疗健康知识融入到对话系统中，本项目设计了针对异构时序数据等电子病历型数据的表示学习模型，能够更准确捕获到电子病历数据中的知识信息。在此基础上，本项目还将医疗信息抽取模块和个性化对话系统相结合，基于历史数据的个性化推荐，构建了个性化的医疗问诊原型系统。.. 项目组培养了4名博士后、5名博士、4名硕士，分别在北京大学、国防科大、腾讯公司、东软集团和华为诺亚方舟实验室等国内外顶尖的科研单位继续从事科研工作。项目组在高水平的国内外期刊和会议上发表带标注的论文共28篇，谷歌学术总被引723篇次。其中CCF A类论文9篇，SCI索引3篇，EI索引20篇，ISTP索引8篇。本项目所取得的成果获得了广泛的认可，多个合作单位都反映本项目所提出的医疗信息抽取模型特别适应于病例数据的特点，并且泛化能力强，在不同的预测任务上都显著提升了效果。