Research On Knowledge Reasoning Based On Fusion Information

With the popularization and development of the Internet,the amount of data grows exponentially.How to manage and apply such a large amount of data has become a big problem.The emergence of knowledge graph provides a new way of management for Internet data.It brings people from matching retrieval to intelligent QA;a retrieval,excavates the potential value of massive data,and solves the prelude for intelligence.With the deepening and development of the concept of knowledge graph,each company has developed its own knowledge graph,and the biggest problem of the current knowledge graph is the lack of knowledge.Previous knowledge reasoning methods have some limitations,such as complex model,poor interpretability and low efficiency.In the classical translation model,the information used only includes entity triples,and the lack of path,entity category and other information in the KG.It is no longer suitable for knowledge graph with complex structures.The core of this thesis is to apply a variety of information form knowledge graph to reasoning.The focus of knowledge reasoning is the completion of knowledge graph.The basis of knowledge graph is entity triples.Therefore,this thesis studies from link prediction and entity classification.For link prediction,path and entity description are introduced to help triples to increase the model's ability to distinguish entity types and extract path relationships.For the large set of paths generated by random walk,the aggregation of the set greatly reduces the number of paths,cuts the cost of the algorithm,and improves the accuracy of path prediction to a certain extent.For entity classification tasks,similar to natural language processing,classifies and marks entities,but the data structure of knowledge graph is different from document,which means the model and method cannot be applied directly.In this thesis the graph convolution neural network is used to extract information,and on this basis,the influence of different entities around the predicted entity is considered.Finally,the output of GCN is used as the representation of the predicted entity,which fully improves the ability of entity classification.In order to verify the effectiveness of methods and models described in this thesis,translating model,DKRL and so on are selected as the comparison,and the same data set as the experiment is used to design the comparison experiment of link prediction and entity classification.Experimental results show that the method proposed performs better than the previous methods,and can help knowledge reasoning of knowledge graph.