Research On Knowledge Graph Completion Algorithm Based On Inductive Logic Program

In order to solve a large number of knowledge management needs brought about by the explosive growth of information data,knowledge graphs came into being.The data of knowledge graph is sparse and incomplete,which is not conducive to its effective use,which also leads to the need for the completion of knowledge graph.Link prediction can use the data of the knowledge graph to complete the missing entities and relationships,and it is the current mainstream knowledge graph completion method.Among the mainstream link prediction methods,the neural symbolic inference method has the advantage of being more robust than the symbolic inference method,and more interpretable than the neural network method.Most of neural symbolic reasoning methods are based on inductive logic programs,and the computational efficiency is often limited by inefficient search algorithms and pruning methods,which are at a low level.This also makes it difficult to apply such methods to large-scale knowledge graphs.In order to improve the computational efficiency of such methods,we studies the knowledge graph completion algorithm of neural backward chain reasoning.In order to improve the computational efficiency of the neural backward chain reasoning knowledge graph completion algorithm and make the pruning process more interpretable,we proposes a neural backward chain reasoning method based On the one hand,the method can capture the hierarchical relationship of knowledge traversal in the neural backward chain reasoning module by using the recurrent neural network? On the other hand,this method can improve the computational efficiency compared with the original neural symbolic reasoning method by statically adaptively dividing the knowledge area.Experiments on public datasets show that the link prediction effect of this method is only slightly lower than that of mainstream neural symbolic reasoning methods under the premise of greatly improving computational efficiency,and the modules of knowledge area division are interpretable.An existing method divides the knowledge traversal area of the neural backward chain reasoning method by adding a neural network layer,which greatly improves the computational efficiency.However,due to its use of the attention mechanism,it has stricter requirements on the size of the embedding.Only when the size of the embedding is much larger than the number of knowledge base predicates can it achieve good results,thus affecting its scope of use.We proposes a neural backward chain inference method based on dynamic knowledge region generation.This method uses a recurrent neural network to re-match and filter the knowledge regions divided by the attention mechanism,thereby reducing the dependence on the attention mechanism and further improving the computational efficiency.Through experiments on public datasets,when the embedding size is close to or smaller than the number of predicates,the method can also obtain relatively close link prediction effects,and the computational efficiency is further improved.At the same time,the knowledge selection module also has a certain degree of interpretability.