基于图模式的知识图谱一致性问题研究

The rapid development of Internet has caused the explosion of data, which provides a rich resource of knowledge. At present, the knowledge graph has got successful application in search, recommendation, query answering systems, and other areas. However, most of knowledge graphs are constructed from multiple sources by using information extraction technologies, which makes there are lots of contradicting knowledge in knowledge graphs. This makes it important to detect inconsistencies in knowledge graphs, and better still, repair the knowledge by fixing the errors without generating new conflicts. This refers to the knowledge consistency, corresponding to data consistency for relational databases. At present, there exists no generally accepted model for knowledge consistency, and moreover, the volume of big data in knowledge graphs and their nature of distribution and evolution have brought huge challenges to cleaning knowledge graphs. In light of this, we aim to pursue researches on knowledge consistency. More specifically, we first establish constraint languages for knowledge consistency based on the notion of graph patterns, to capture errors in real-life knowledge in knowledge graphs. Then we study the mining algorithms to discover consistency constraints from sample knowledge, and develop techniques to detect violations of consistency constraints, and fix the errors and make the knowledge graph consistent.

信息社会的快速发展引发了数据规模的爆炸式增长，为知识图谱的构建提供了丰富的资源，使其初步成功应用到了搜索、推荐、问答系统等领域。然而，由于信息抽取等技术的运用及网络数据源的不可靠，知识图谱往往带有冲突的知识而出错，需要快速、准确的错误发现方法，以及具有可靠性保证的错误实时更正方法，即保证更正过程中不会产生新的错误，这被称为知识图谱的一致性问题。目前，知识图谱作为图数据，缺少有效的模型来刻画其一致性需求。同时，知识图谱的规模巨大、分布性及演化性给其一致性保证带来了巨大的挑战。因此，本项目研究知识图谱的一致性问题，基于图模式建模知识图谱的一致性约束，以刻画产生冲突知识的语义错误，设计一致性约束的自动挖掘方法以产生有效的一致性约束，提出高效的不一致性的自动检测算法以快速发现冲突知识，建立可靠性保证的不一致性的自动修复算法，以实时修复知识图谱中的冲突知识，并保证更正后的知识图谱不会产生新的错误。

知识图谱被广泛应用于搜索、推荐、问答系统等应用领域，但通常存在冲突的知识，降低了用户的满意度。因此，如何刻画知识图谱中的冲突知识，并对其快速、准确的发现和修复从而保证知识图谱的一致性对知识图谱得到有效应用起到重要的作用。同时，知识图谱的规模巨大、分布性及演化性给其一致性保证带来了巨大的挑战。本项目对知识图谱的一致性问题，开展对一致性约束模型、一致性约束发现方法、不一致性检测方法和修复方法的研究。.在本项目执行期间，项目组建立了基于图模式匹配的一致性约束模型和其公理化系统，对其可满足性问题、逻辑蕴含问题及验证等问题进行了复杂性分析，设计了一致性约束的并行发现算法；提出了不一致性增量检测方法以快速发现冲突知识；提出了正确性保证的不一致性的并行和增量修复算法，以快速修复知识图谱中的冲突知识，并保证修复后的知识图谱不会产生新的错误。