Research On Knowledge Graph Question Answering Based On Representation Learning

With the rapid development of Internet and artificial intelligence technology,traditional search engines have been unable to quickly obtain accurate information.In the era of big data explosion,knowledge graph came into being.It will be faster and more accurate to obtain relevant answers through knowledge graph.However,the current technology cannot build a very perfect knowledge graph.The knowledge graph constructed manually has a small amount of storage and is time-consuming.Therefore,the existing knowledge graph still faces incomplete problems,which will lead to low accuracy of question answering based on knowledge graph.With the development of representation learning technology,some studies have applied knowledge graph embedding technology to the field of knowledge graph question answering,but this method ignores the influence of super nodes in knowledge graph.Therefore,this paper discusses how to eliminate the influence of super nodes in the process of completing knowledge graph,and proposes a knowledge graph question answering method based on path traversal.The main research contents and work of this paper are as follows.(1)Research on knowledge graph completion based on deep learning technology.Aiming at the problems of incomplete and super nodes in knowledge graph,this paper studies the performance of knowledge graph completion through path traversal in the embedding space of knowledge graph.The path traversal of the knowledge graph embedding space mainly includes two steps : path search and path reasoning.The translation vector of each node is obtained by path search.The obtained translation vector is analyzed by path reasoning to obtain the probability of the relationship between the head entity and the target entity.Finally,a large number of comparative experiments were carried out.The experimental results show that the MAP on the NELL-995 dataset reaches 0.934,the MAP on the FB15K-237 dataset reaches 0.864,and the MAP on the WN18 RR dataset reaches 0.848,which are higher than the baseline model.Therefore,the performance of path traversal prediction links in the knowledge graph embedding space is better.(2)Research on problem entity recognition based on self-attention mechanism.Based on BiLSTM-CRF,this paper proposes a BiLSTM-CRF problem entity recognition model based on self-attention mechanism.This paper uses self-attention mechanism to filter and encode words in sentences.First,the model uses GloVe to embed words into the problem,which are then sent to the BiLSTM layer,the self-attention layer,and the CRF layer in turn.The experimental results show that the F1 value of BiLSTM-CRF with self-attention mechanism on entity recognition reaches89.27%,which is 0.37% higher than that of BiLSTM-CRF directly,showing better results.(3)Research on multi-hop knowledge graph question answering based on path traversal.This paper applies the method of traversing knowledge graph embedding space to multi-hop knowledge graph question answering,and proposes TrEKBQA method.The method mainly includes three modules : knowledge graph completion,multi-hop knowledge graph question and answer selection.The knowledge graph is completed by traversing the embedding space of the knowledge graph,and then the question is answered.Finally,the appropriate answer is selected according to the value of hits@1.The experimental results show that the hits@1 of TrEKBQA on the original MetaQA-1hop,MetaQA-2hop,MetaQA-3hop and WebQSP datasets reached 98.3%,99.4%,96.1% and 73.2%,respectively.The hits@1 on the incomplete MetaQA-1hop,MetaQA-2hop,MetaQA-3hop and WebQSP datasets reached 85.1%,92.3%,70.9% and 54.7%,respectively,which were higher than the baseline model.These results show the effectiveness of TrEKBQA in knowledge graph question answering.