Research On Reasoning Algorithms For Combination Of Semantic Paths In Knowledge Graph

With the massive growth of data on the Internet and the development of artificial intelligence in recent years,the ability to process a variety of heterogeneous data is also constantly improved,Knowledge and its relationship are gradually valued by industry.The development of the knowledge graph provides computers with understandable structured semantic information,which is an important driving force for the development of intelligent human society.Its establishment makes the data on the Internet have better semantic processing capabilities,but the knowledge graph is still under the development,which is far from reaching the degree of perfection.Although the knowledge graph contains hundreds of millions of facts,but it has not completely covered some knowledge that is not commonly used.Meanwhile,some knowledge that can be obtained by simple reasoning have not been identified.Therefore,it is of great significance to complete the knowledge graph.At present,most of the work is focused on the completion of the direct relationship between the entity pairs,but this article focuses on the completion of some factual relationships that exist in short paths.For the reasoning problem of the implicit relationship in the current knowledge graph,the work of this article is as follows:In the past,the process of knowledge reasoning was mostly based on graph structure,which has high calculation cost and poor extensibility.This article is inspired by the fact that knowledge represent learning.Using the translation model proposed by Bordes A et al,the knowledge graph is embedded in low-dimensional vector space for calculation,which improves the calculation efficiency and scalability.In order to mine the information on the path in the knowledge graph,it is necessary to obtain some reachable path information existing between the two entities.In this article,a path discovery model based on reinforcement learning is proposed,which uses the interaction process between agent and environment,and the model learn to find the path between two entities under the mechanism of the reward function.This will pave the way for using the path information to infer relationships.After obtaining the path information between two entities,in order to make better use of the entities and relationships on the path,this paper uses the RNN model to deal with the characteristics of sequence problems of any length,and it uses the obtained entities and relations as the input of RNN.After iterative calculation,it will finally get a result vector combined with the path information,then the result vector and the target relationship vector are similarly calculated,and finally the relationship is determined by the similarity value.Finally,in order to verify the effectiveness of the proposed algorithm,experiments and analysis of the algorithm are carried out,and it shows that the algorithm effectively improves the accuracy and precision of the results.