Research On Knowledge Graph Completion Algorithm Based On Multi-hop Relation Question Answering

Because of its highly structured form,knowledge graph has been widely used in many AI fields such as question answering system,intelligent recommendation and intelligent customer service.Although these knowledge graphs contain a large amount of knowledge information and perform well in related applications,there is still the problem of sparsity.How to extract external knowledge to supplement the knowledge graph or complete it on the existing knowledge graph has become an important research hotspot.Representation learning method based on knowledge graph embedding shows good results in its completion task.Knowledge representation learning is an efficient means to realize knowledge graph completion.It represents the entities and relations in the knowledge graph as low-dimensional dense vectors,and inferences the potential semantic relations in the way of vector computation.However,these algorithms can only deal with the case that there is only a first-order relationship between entities,and they lack the expression ability in the case involving multiple relationships between entities.Therefore,this paper aims at the potential fact completion in the multi-relational path in the knowledge graph,and proposes a combination model of the knowledge graph completion algorithm based on multi-hop relational question answering,RL+ Rotate.The work contents of this paper are as follows:(1)In the previous multi-hop question and answer,the method of reinforcement learning was adopted.According to the query,it went upstream in the knowledge graph and stopped when it reached the target node.However,in the training process,the normal samples in the test set will be considered as errors,and then the Agent will be rewarded with a low return,which will lead to a decrease in the accuracy of the test results.Inspired by the knowledge graph embedding technology,this paper uses the knowledge graph embedding model to pre-train the triples,so that each triplet has an expected reward,so as to effectively avoid the influence of false negative examples on the test results during training.(2)Based on the pre-training of the knowledge graph,the final answer is obtained through the knowledge graph path reasoning process.In this case,the answer and query elements form a triplet that is simultaneously validated as a fact triplet.To some extent,it plays the purpose of knowledge graph completion.(3)The model in this paper is tested on the relevant knowledge map data set.Compared with the existing similar algorithm models,it is found that the test results of the combined model proposed in this paper are improved.It is shown that the combinatorial model is effective in the task of knowledge graph completion for multi-hop relationship question answering.