Research On Complex Question Generation Over Knowledge Graph With Deep Learning

Question Generation over Knowledge Graph(KGQG)aims to generate corresponding questions according to the input Knowledge Graph(KG)subgraph and given answer entities,which can be used to automatically construct large-scale question answering datasets to support the research of Question Answering over Knowledge Base(KBQA),and has been widely concerned by researchers in recent years.This thesis studies the complex question scenarios in the task of KGQG.Based on a more general setting,the input KG subgraph can contain multiple triples,and the generated question contains complex multi-hop relationships.Existing methods based on Deep Learning are generally convert input into sequence for encoding,which cannot effectively encode non-Euclidean spatial structure data such as the input KG subgraphs in complex question scenarios,and some methods using Graph Neural Networks(GNN)cannot effectively learn multi-hop dependency information between nodes of KG subgraphs.At the same time,existing methods have no reasonable way to constrain the question generation process and ensure the complexity of generating questions.To tackle the problem mentioned above,this thesis proposes a complex KGQG model based on GNN.This model uses GNN to effectively encode the input KG subgraph,and the nodes aggregate their own multi-hop node information to update the vector representation.Combined with the distance information between different nodes,the multi-hop dependency relationship between nodes is learned and the structure information of the input KG subgraph is effectively utilized.This thesis also proposes a complex KGQG model based on joint learning,which constrains the question generation process and ensures that the generated question contains the expected complex relationship through the joint learning of the relationship extraction task and the KGQG task.By introducing joint tasks,the global relation node vector representation of the output of the KGQG model is constrained,and the question generation process is completed under the guidance of global relations,so that the model can accurately infer the required information in the process of generating question and effectively improve the quality of generating questions.At the same time,this thesis design more reasonable and effective joint learning methods to improve the performance of joint learning.The main contributions of this thesis are as follows:1)A complex KGQG model based on Bi-MHDGT networek is proposed.In this model,the Bi-MHDGT networek proposed in this thesis is used to encode the input subgraph,considering the bidirectional information of nodes,and combining with the structural information of subgraph,the multi-hop dependency relationship between nodes is effectively learned,thus improving the encoding ability of complex subgraph.2)In this thesis,a complex KGQG model based on Joint-Turbo learning is proposed.In this model,the relationship extraction task and the KGQG task are used for joint learning,and the global relationship vector representation is constrained to guide the question generation process,so as to ensure that the question containing the expected complex relationship is generated.In order to improve the performance of joint learning,this thesis proposes a turbo joint learning method.3)Experiments are designed on two complex question datasets which are widely used in the task of KGQG.Com-pared with the existing mainstream methods,the effectiveness of this method is verified,and the influence of the important parts of this method on the final results is analyzed through ablation experiments.Experimental results show that the proposed method can effectively improve the quality of generating problems,and generate question with expected complex relationships more accurately.To sum up,this thesis studies the method of KGQG based on deep learning,and proposes a model of comple KGQG based on Bi-MHDGT network and a model of complex KGQG based on Joint-Turbo learning,which can effectively realize the generation of complex question by improving the encoding ability of complex subgraphs and restricting the question generation process in a reasonable way.