| The knowledge base(KB)realizes the mapping of objective facts in the real world.The Knowledge Base Question Answering(KBQA)can effectively retrieve the massive factual data,so it is widely used in artificial intelligence and other fields.For the KBQA,users can quickly and accurately query answers.In order to better meet the needs of real-world scenarios,the KBQA often requires complex and multi-hop reasoning.But for now,it is still a challenge to improve the accuracy of complex multi-hop questionanswering tasks in the study of KBQA.First,the question-answering reasoning model has insufficient performance when faced with complex multi-hop questions.Second,the reasoning ability of question-answering reasoning models relies on the quality of question-specific subgraphs.The trade-off between question-specific subgraph size and answer coverage rate affects the accuracy of the question-answering reasoning model.To response the above problems,this dissertation first constructs an implicit reasoning model based on the semantic information of relations and neural network structure to improve the ability of KBQA to solve complex multi-hop questions.Secondly,this dissertation constructs a subgraph retrieval model,which generates simplified subgraphs through bidirectional path reasoning,thereby improving the accuracy of KBQA.The main work and contributions of this dissertation are as follows:(1)Aiming at the insufficient performance of the question-answering reasoning model on complex multi-hop questions,an implicit reasoning model based on relational distribution is proposed.For the KBQA task,the path of reasoning is a deformation of natural language questions,and the intermediate entity information in the reasoning path is irrelevant to natural language questions and will bring semantic noise.According to this,this dissertation focuses the reasoning of KBQA on the semantic information of relations in the reasoning path.Furthermore,the model performs implicit reasoning using a neural network architecture for aggregating semantic information of relations.Finally,this dissertation conducts experimental verification of the model on the current popular question-answering dataset.The experimental results show that the implicit reasoning model based on the relational distribution achieves better results than the baseline model.(2)Aiming at the difficult trade-off between question-specific subgraph size and answer coverage rate,a subgraph retrieval model based on bidirectional path reasoning is proposed.For the KBQA task,a subgraph with a low degree of simplification contains more noisy nodes,and a subgraph that is too simplistic loses the answer node.According to this,this dissertation first preliminarily filters relations in the subgraph,and then determines the correlation between reasoning paths and the question from the perspective of forward and reverse,so as to retrieve the subgraph and realize the simplification of the subgraph.The experimental results show that the subgraphs retrieved by the model are better than the baseline model,and after the subgraphs are fed back to the reasoning model for verification,the accuracy of the question-answering reasoning model is also improved. |
