Research On Rule Learning Of Large-scale Knowledge Graphs

With the unprecedented development of Artificial Intelligence,knowledge graphs(KGs),as the new generation of knowledge bases,have received significant attention in both academia and industry.The name “knowledge graphs was coined by Google in 2012”.KGs can be used to describe the various entities and concepts existing in the real world and the relationship between them.The knowledge graphs can make up for the description ability of Machine Learning and Deep Learning algorithms.The representations such as rules are explicit knowledge compared with neural networks.As a result,link prediction based on automatically learnt rules can provide human understandable explanations.In this dissertation,we first investigate the problem of rule learning by employing embedding in representation learning and then apply learnt rules in link prediction for KGs.We first propose a hierarchical sampling algorithm.When the sampling process is used with rule search algorithms,the rule search complexity can be significantly reduced.Improved vector embedding score functions are used to reduce the amount of intermediate calculation.We combine the calculation process of the rule evaluation with the confidence of the quality assessment,and propose optimization measures for the evaluation of the rule head.Moreover,the link prediction by rule-based reasoning is presented.The learned rules are further semantically extended by generating high quality rules,which is realized by mining semantic similarities and replacing rule atoms with close relationships.Thus,the scalability and efficiency of rule learning is achieved by our new techniques.As a result,a system R-Linker is developed for link prediction in large KGs.Our system outperforms some existing methods for link prediction in terms of accuracy.