A Research Of Relational Inference Algorithm Based On Knowledge Graph

The success of Google Knowledge Vault,Microsoft Satori,IBM Watson and other knowledge graph projects have aroused widespread attention from the industry and academia to knowledge graph technology.Knowledge graph use entities and relations to represent objects in the real world and relations which between different objects.It provides the ability to organize,manage and understand the vast amount of unstructured information on the internet,and the representation of information is closer to the human cognitive world.Therefore,knowledge graph is the core foundation in the fields of Natural Language Processing,knowledge engineering and intelligent information retrieval.It is also the cornerstone of artificial intelligence.However,due to the limitation of information extraction technology and the incompleteness of existing data sources,the incompleteness of knowledge has become the main bottleneck,which restricts the application and development of knowledge graph.The relational inference technology provides an effective solution for the above problems,and is the current main technical measure for knowledge graph population and knowledge quality assessment.The main idea of the relational inference technology is to utilize the existing knowledge in the knowledge graph and automatically deduce the missing relation between entity pairs,which has become one of the core technologies to promote the development of the knowledge graph.This thesis study the relational inference technology in knowledge graph.Through the comprehensive investigation of the state-of-the-art related work at home and abroad,the models proposed in recent years can be divided into three categories according to modeling methods.This thesis focuses on the study of two kinds of methods,the main research contents are summarized as follows:1.Studying representation learning based relational inference methods: By analyzing the existing representation learning based relational inference models,we find the general potential problem contained in their basic assumptions,that is,ignoring the semantic diversity of entities and relations.Accordingly,a nonlinear transformation modeling method is proposed to solve the problem of semantic resolution in the representation learning based models.Based on the above idea,this thesis designed and implemented the unified weighted model and the independent weighted model.Experiments show that the proposedmodels are significantly outperform related works.2.Studying logic rules based relational inference methods: The study conduct by this thesis shows that there are two problems in PRA and its related algorithms.Firstly,the algorithm extracts the relational path features through random sampling,which improves the computational efficiency while sacrifice the utilization of existing information in knowledge graph.Secondly,using the supervised learning method to establish the relational inference model,the effectiveness of the model depends on the training data,especially affected by data sparsity.Accordingly,the bidirectional semantics hypothesis and the inferential of relational-specific graph hypothesis are proposed,and the Two-tier Random Walk Algorithm is designed and implemented.The experiments on open datasets verify the rationality of the above assumptions and the effectiveness of the proposed algorithm.