Reasearch Of Representation Learning Methods For Knowledge Graph Completion

Knowledge graphs aim to organize human knowledge in a structured way and are an important support for artificial intelligence research.At present,most knowledge graphs still have a large number of implicit relationships between entities that have not been explored leading to incomplete graphs,so knowledge graph completion becomes an urgent problem to be solved.Knowledge representation learning is one of the main methods for knowledge graph completion,and this thesis conducts an in-depth study on knowledge representation methods for knowledge map completion based on uncertain knowledge graphs.For the existing uncertain knowledge graph representation learning methods still have the problems of insufficient utilization of auxiliary information and false-negative samples will appear in training,the main research contents of this thesis are as follows:(1)Research on semantic augmentation-based box-embedding knowledge representation learning method.In this thesis,we propose a semantic augmentation-based box-embedding knowledge representation learning model to address the shortcomings of existing knowledge representation models that ignore auxiliary information.Firstly,the word sequences connected by head and tail entities and relations are encoded as word embedding sequences,and then the semantic features with the head and tail association information with confidence are mapped to the box embedding learning representation space using the attention mechanism and Bi-LSTM network structure as the initialized features of the box embedding,and finally the representation learning of triples is performed by the box embedding.Experiments show that the proposed model in this thesis has improved its effectiveness over other representation learning models on public datasets.(2)A study on the method of knowledge graph representation learning model based on semi-supervised learning.To address the problem that existing models cause a large number of false negative samples in training,this thesis proposes a knowledge graph representation learning model based on semi-supervised learning,which not only treats the samples that do not exist in the graph as negative samples,but also re-evaluates their confidence levels to alleviate the problem of false negative sample problem.The experiments validate the practicality and effectiveness of the method.(3)A knowledge graph completion system is designed and implemented.The system integrates the semantic augmentation-based box-embedding knowledge representation learning method and the semi-supervised learning-based knowledge graph representation learning model method to realize the knowledge graph completion function and verify the practicality of the model proposed in this thesis.The system realizes the functions of knowledge query,knowledge graph completion,confidence evaluation and data management.