Building and Querying Knowledge Graphs from Text Corpora

知识图（Knowledge Graphs, KGs）是一种具有灵活数据模型的知识库，可以无缝地表示来自传统数据库的结构化数据和来自文本的半结构化信息。图中的节点是真实世界的实体或其属性，而边将实体连接到属性或其他相关实体。KG有一个本体，描述了类型层次结构以及KG中关系的域和范围。在过去的十年中，建立了几个大型和通用的KGs，并发现了许多重要的应用：使语义搜索、问题回答和远程监督的深度神经方法接近人类水平的性能。KGs还用于在关联开放数据（LOD）云上共享知识。虽然KGs的理论基础很好理解，并且有管理KGs的可靠系统，但我们缺乏从已有数据集创建相互关联的KGs的原则过程。

Knowledge Graphs (KGs) are knowledge bases with a flexible data model that seamlessly represent structured data from traditional databases and semistructured information derived from text. Nodes in the graph are real world entities or their properties while edges connect entities to properties or to other related entities. KGs have an ontology, describing a type hierarchy and the domain and range for the relations in the KG. In the past decade several large-scale and generic KGs were built and found many important applications: enabling semantic search, question answering, and distant-supervision for deep neural methods nearing human-level performance. KGs are also used for sharing knowledge on the Linked Open Data (LOD) cloud. While the theoretical underpinnings of KGs are well understood and there are solid systems for managing KGs, we lack a principled process for creating interlinked KGs from already existing datasets. This research program will contribute principled algorithms and system for building factual and interlinked KGs from an existing corpus of semistructured documents (mixing text, tables and lists) and a reference ontology for interlinking purposes as well as algorithms for translating questions in natural language into structured queries that can be answered from the Kgs. The research carried out through this Discovery Grant will leverage the state-of-the-art in information extraction from text, machine learning, and scale-out data management techniques on shared-nothing clusters. The tools developed through this Discovery Grant will allow domain experts to extract KGs from existing datasets so that they can share that knowledge of make sense of it via structured queries. Thus, these tools will contribute to decision making, which in the modern knowledge economy we live in requires making sense of heterogeneous data coming from structured databases and an ever increasing volume of text (email, legislation, technical literature, et.c). Moreover, the HQP trained through this program will acquire skill that are currently in high demand in industry and will remain so for the foreseeable future. Finally, the tools developed through this program, by virtue of being open and cloud-based, will allow researchers and educators, across disciplines, to experiment with and contribute to the development of KGs in their domain and the training of HQP of their own.