Research Of Knowledge Inference Method Based On Low Dimensional Vector Space

With the rapid development of Internet technology and the explosively increasing amount of data,people are showing increasing demand on the precision and the intelligence of searching.The rise of knowledge graphs opens the prelude of intelligent information retrieval,from conventional string matching to intelligent understanding.Knowledge graphs can be used for deep text mining and accurate question answering,thus is one of the core technologies of artificial intelligence.However,existing knowledge graphs are still facing the serious problem of lacking knowledge.To tackle this challenge,knowledge representation learning has been extensively investigated,which can promote knowledge acquisition,integration,reasoning and knowledge graph completion.Traditional discrete symbolic knowledge representation has the problems of high computational complexity and low scalability to large data.Instead,distributed knowledge representation based on low dimensional vector space has become a new trend for solving such problems.However,most existing knowledge representation learning approaches only utilize the structure of triplets defined in knowledge graphs,regardless of the entity descriptions and categories which conveys rich semantic information of entities.Therefore,those approaches demonstrate poor performance on large-scale spares knowledge graphs.Driven by these observations,in this thesis,we will mainly investigate effective knowledge representation learning approaches that are capable of incorporating multi-source information.Besides,we will explore knowledge inference using knowledge vector representations in order for acquisition new knowledge.To address the problem that current knowledge representations cannot incorporate complete semantic information contained in entity descriptions,in this paper,we proposed a new representation learning algorithm,named RLCD,for incorporating entity descriptions,which is based on the TRANS* models and deep learning models including DOC2 VEC and LSTM.The earliest approach to knowledge representation learning with entity description,named DKRL,takes as input the triplets of the knowledge graph with part of the high-frequency words in entity descriptions,which fails to capture the complete semantic information within entity descriptions.The RLCD algorithm proposed here utilizes document embedding models to obtain document vector representation which is directly fed as the input of our model,in order to minimize the information loss.In addition,by incorporating word order of sentences,our obtained knowledge representa tion not only includes the semantics of triplets in knowledge graph,but also the complete semantics in entity descriptions.Our experimental results on FREEBASE and PHONECARD datasets show that,compared with DKRL,our algorithm achieves significantly better link prediction betters while being faster.The obtained knowledge representations will be more beneficial for knowledge graph completion and knowledge inference.Another important and challenging research problem is using knowledge graphs for knowledge inference,in order to acquire new knowledge.Traditional knowledge inference approaches typically use inference strategies derived from logic forms and association rules,which have the disadvantages of low coverage,low inference speed,low scalability and incapability of representing uncertainty.On the contrary,knowledge inference based on distributed representation can be much more efficient,by using vector computations instead of graph traversal and search as in traditional knowledge inference approaches.In this paper,we will learn distributed knowledge representations that encode both relations paths and entity descriptions by using knowledge representation learning models and deep learning,to discover the indirect relations between entities entailed in their relation path,which further benefits the inference of complicated relations.