Research On Question Decomposition Based Complex Question Answering Over Knowledge Graph

Question Answering over Knowledge Graph is one of the most promising research fields in natural language processing.One mainstream method of KGQA is Semantic Parsing-based methods,which parse the question into formal query and execute it against KGs for finding the answers.These methods have achieved good results in solving simple question,but encountering serval challenges when handling complex questions.Specially,the challenges are(1)understanding complex question semantics;(2)parsing complex queries;(3)grounding with large search space.One breaking point is enhancing the semantic feature learning of complex question.This thesis proposes the effective question semantic feature learning methods to overcome existing challenges.The main contributions of this thesis are listed below.(1)Proposing History-Attention based Formal Query Generation method(HAG)that based on Formal Query Staged-Generation Framework.Firstly,this thesis designs Formal Query Staged-Generation Framework whose modules reflect mainstream design concept.Secondly,we find existing works are devoted to increase the importance of key information,but ignore the interference of redundant information when learning question semantic feature.Based on this,History-Attention based Formal Query Generation method is proposed and performs well in two benchmark datasets.(2)Proposing Question Decomposition based Formal Query Generation method(QDG).Applying Question Decomposition into KGQA has two main challenge:(1)lacking of labeled data for training;(2)determining the answer order of sub-questions.For lacking of labeled data,this thesis obtains rough labels from attention weights of HAG and trains question decomposer via weakly supervision training.For determining the answer order of sub-questions,in this thesis,the History-based Attention combined with Self-Attention is used to indicate question decomposition model.Extensive experiments on two benchmark datasets have demonstrated the effectiveness of our approach on decomposing question and handling complex questions.(3)Designing and implementing the Knowledge Graph based Question Answering system called QDG-KGQA.Graphic User Interface of QDG-KGQA provides users with intuitive interaction.In addition to the answers,the system shows the sub-question generated during the answer process,which enhanced reliability and interpretability of the system.The significance of this work is that exploring the feasible solutions of enhancing question semantic feature learning.Specially,this thesis aims to verify the effectiveness of question decomposition for KGQA,especially for handling complex question.Our work proposes the feasible solutions to overcome existing challenges which have great reference value for future work.