Research And Implementation Of Graph-based Relation Inferring Algorithm

With the rapid development of information technology and the Internet,the traditional information retrieval method which based on text has been difficult to meet the needs of users.People have the urgent need for a more efficient information processing method.The research of Knowledge graph is generated and developed to deal with the problem,which extract the structured information from the non-structured text,and ultimately build a huge knowledge network.The relevant application based on this network provides users with faster and more accurate services.Relational reasoning technology is one of the key technologies in the construction of knowledge graph.It can use the knowledge already existing in the knowledge graph to obtain new knowledge through reasoning and prediction.The technology is mainly used for the expansion of knowledge graph,and the applications like Q & A system.Based on the research in the field of relational reasoning at home and abroad,this thesis proposes a new graph-based relational reasoning algorithm GBRI.The algorithm mainly includes two parts: global relational reasoning module and local relational reasoning module.The two modules use the connectivity structure between different relationship types and the structural features in the single relation to make the reasoning prediction,and get the final relational reasoning result.The main contributions of this thesis are as follows:(1)A new relational preprocessing method is proposed,which combines the synonymous relations in the knowledge graph,reduces the dimension of the relational feature space.It makes the training data in the model learning process more abundant,alleviates the data Sparseness problem.(2)A global relational reasoning algorithm is proposed.The algorithm maps the knowledge graph with an undirected graph and constructs the "entity-relation" graph.Then,for each relation contained in the knowledge graph,the algorithm uses Logistic regression to train the prediction model of each kind of relationship.Experiments show that use undirected graph model can effectively increase the number of feature paths that can be obtained,and finally improve the prediction performance of the algorithm.(3)A local relational reasoning algorithm is proposed,which extracts the relational subgraphs of each relation from the "entity-relation" graph and calculates the transition probability between different entities in the subgraph,and then obtains the prediction result of the local relational reasoning model.The Experiments show that the proposed algorithm can effectively improve the prediction performance of many-to-many relationship data.(4)In this thesis,we combine the global relational reasoning algorithm and local relational reasoning algorithm to obtain the final GBRI algorithm.And the effectiveness of the GBRI algorithm is proved by comparing the performance of the GBRI algorithm with other representative work in three open source datasets: WN18,FB15 k and FB14 k.