Research On Knowledge Graph Reasoning Method Based On Deep Learning

The knowledge graph uses a structured way to store entity information and the relationship between entities,which can process and utilize information quickly and efficiently.At present,knowledge graphs have made great breakthroughs in the fields of intelligent search and intelligent question answering,but there are still problems such as missing and sparse data.How to use existing triples to infer potential fact triples to supplement knowledge graphs is an urgent need.solved problem.In recent years,deep learning has made breakthroughs in many fields,but in knowledge reasoning application scenarios,the knowledge reasoning process will become unexplainable.How to improve the interpretability of deep learning and integrate it into knowledge reasoning is still a problem.In-depth research is required.In response to the above problems,this paper conducts research from two aspects: deep learning network and knowledge reasoning method.Firstly,aiming at the problems of excessive storage of Transformer network parameters and low computational efficiency of sparse full attention in deep learning,an efficient neural network HET model based on hash function is constructed.Axial position encoding is used to represent position information;a reversible residual network is introduced to complete backpropagation and update parameters without storing the activation of the intermediate layer;the full attention calculation is transformed into a hash-based approximate nearest neighbor retrieval problem,which is obtained by random projection hash value to complete the sparse attention calculation.Secondly,aiming at the problems of high instantiation complexity of rule-based methods and lack of interpretability of neural network-based methods in knowledge reasoning methods,a HETIL knowledge reasoning model that mixes logical rules and neural networks is constructed.First-order logic rules are used to extract paths between triplets as rules,and logical symbols are used to construct a multi-layer rule space;rule feature information is extracted through an efficient neural network HET model,and the rule scores are calculated by instantiating.Finally,machine translation experiments are carried out on the public dataset CCMT2018 English-Chinese translation,and the BLEU score is used as the evaluation index to verify the effectiveness of the performance of the HET model;knowledge reasoning experiments are carried out on the public datasets WN and FB series,using MMR,Hits@10 and single query time consumption are used as evaluation indicators,and a comparative experiment with the classic knowledge reasoning model is carried out to verify the effectiveness of the HETIL model.