Research On Knowledge Reasoning Based On Knowledge Graph

In recent years,with the rapid development of the Internet,data and information are growing explosively.How to effectively process data on the Internet has attracted widespread attention.Based on this,Google puts forward the concept of knowledge graph.The knowledge graph is essentially a multi-relationship network,which uses a structured way to store the knowledge system of the corresponding relationships between entities in the real world.Knowledge graphs are widely used in intelligent question answering,recommendation,information retrieval and other fields,and become an indispensable data support in the field of artificial intelligence.It is of great significance to promote the development of artificial intelligence.However,due to the current information extraction technology,many knowledge graphs have the problem of missing data,and the lack of data also restricts the development and application of knowledge graphs.Knowledge reasoning technology is to use the existing knowledge in the knowledge graph to make inference predictions to obtain the missing hidden knowledge in the knowledge graph.It is an important method to automatically complete the knowledge graph by using a computer.This thesis studies knowledge inference technology based on knowledge graphs,introduces and analyzes the traditional translation distance model in detail,and proposes new knowledge inference algorithms from the perspectives of fusing external information and nonlinear modeling.The research content is as follows:(1)Propose a knowledge reasoning model that integrates entity description information in the knowledge graph(KRLEDT).In order to obtain all the semantic information of the entity description information,this thesis uses the Doc2 Vec model to pre-train the entity description information to obtain the embedded representation of the sentence granularity of the entity description information.Then use the Transformer(Encoder)to further extract the word order information in the entity description information of the sentence granularity,and finally combine the Trans E model to learn the triplet structure information,use the translation invariance assumption to model,and add the structure information and the potential energy function.Cross-item representation of entity description information.The link prediction task was performed on the public data set FB15 K.KRLEDT model reached 71 and 75.3%(Filt.)on the Mean Rank and Hits@10 indicators respectively.On the triple classification task,the accuracy rate Reach88.2%,which is better than the comparison algorithms.(2)Propose a knowledge reasoning model based on triple structure information and introduce nonlinear modeling methods(KRL-LSTM).In order to improve the training efficiency of the model,first still use the Trans E model to pre-train the entity and relationship vectors,then treat the triples in the knowledge graph as short texts,use BiLSTM to extract the structure information of the triples,and then use The multi-layer perceptron mapping obtains the score corresponding to the triplet.Experiments were conducted on the public data sets FB15 K and WN18.In the link prediction task,Mean Rank and Hits@10 indicators were used for evaluation KRL-LSTM model reached 48 and 84.3%(Filt.)on FB15 K respectively.On the WN18 data set,it reached 331 and 95.1%(Filt.)respectively.On the triple classification task,the accuracy of the proposed algorithm on the FB15 K data set reached 91.4%,which was partly better than the comparison algorithms.