Research On Knowledge Graph Completion Algorithm Based On Knowledge Representation Learning

Knowledge graph is a database of graph structure,which describes the concept of things in the objective world and the connections between them in the form of triples.It has been widely used in tasks such as recommendation systems,question answering systems,and search systems.However,there are a lot of missing entities or relations in the current knowledge graph.For example,there are about 3 million human entities in the Freebase database,71% of the entities have no birthplace,and 91% of them have no information about their education.These deficiencies could seriously affect the application of knowledge graphs and the accuracy of downstream tasks.Therefore,the task of knowledge graph completion has important research significance.Knowledge representation learning is to map entities or relations into low-dimensional dense vectors and retain the original semantic structure information.It is widely used in knowledge graph completion tasks due to its computational efficiency and alleviating data sparseness.This thesis mainly studies the knowledge graph completion algorithm based on knowledge representation learning.The research work of this thesis is as follows:(1)Existing knowledge representation learning methods that integrate path information rarely filter paths,and the weight of each path is fixed during the training process.To solve this problem,a knowledge representation learning method based on path selection and adaptive weights(Path-selection based adaptive Trans E,PSATrans E)is proposed.Path selection can remove redundant semantic information.The adaptive weights module enables the model to focus on the underfitted path part during each iteration.The knowledge graph completion task shows that both path selection module and adaptive weights module play a role,effectively improving the quality of relation and entity vector representation,and improving the accuracy of completion.(2)Aiming at the current situation that most of the existing knowledge representation models are manually designed,automated machine learning is introduced into knowledge representation learning,and a knowledge representation learning method based on automated machine learning(Complete Neural Architecture Search based Knowledge Graph Embedding,CNASE)is proposed.After amplifying operators of different granularities,the search space covers each method branch.Different model structures can be searched according to the characteristics of different datasets.The experiment of knowledge graph completion shows that it can produce vector representations with richer semantic information than the artificially designed models,and in most cases the effect is better than the artificially designed models.(3)Through the induction and abstraction of the knowledge representation model based on semantic matching,it is found that it can be summarized as a representation method based on a single-layer relation matrix.Accordingly,a representation learning method based on multi-layer relation matrix(Multi-layer Relation Matrix based Knowledge Graph Embedding,MLRME)is proposed.The algorithm expands the relation matrix from the single-layer space of the traditional method to the multi-layer space.The theoretical analysis proves the symmetry,asymmetry,anti-symmetric and anti-relational properties of the multilayer matrix,indicating that it has excellent properties.The expressive ability,and demonstrated superiority in the task of knowledge graph completion.