Research And Implementation Of Aggregate Complex Question Answering Method Based On Coupling Neural Network And Symbol Execution

Knowledge-based question answering system(KBQA)is becoming more and more widely used,and the questions are more complicated.It usually contains multiple query intents and requires logical,quantitative,and comparative inference operations.Semantic parsing is a practical method,which aims to transform natural language queries into executable logic forms(LF),and then obtain final answers through symbolic execution.Semantic parsing usually requires the gold logic form labeled by experts.However,in practical applications,such gold logic forms are expensive to obtain.The weakly supervised semantic analysis model only uses question and answer pairs for training,preferentially searches for feasible logical forms,and then obtains supervised information by comparing the answers with the execution results.However,in the search process,there may be many wrong logics with correct answers.It is difficult for the model to distinguish between false logic forms,and it is difficult to search for a completely correct logic form with higher feedback,so that the accuracy cannot be improved.In this thesis,to tackle the problem mentioned above,we combine neural networks and symbolic execution methods to construct a weakly supervised semantic analysis model,so that the question-answering system can complete the training without logical form annotation.In addition,it attempts to combine active learning strategies to enable the system to actively select a subset of Q&A examples,obtain additional manual annotations as supervisory information,break the optimization-blocked state,and further improve the performance of the Q&A system.The main contributions of this thesis are as follows:(1)Proposing a weakly supervised semantic analysis model which convert the natural language problem into a logical form sequence through the neural network model,and then execute the modularized complex problem by modular symbolic execution.And only use the answer as supervision information for initial training.(2)Proposing a method combining active learning strategies.Based on training a weakly supervised semantic parsing model,heuristically select a part of the data to obtain additional annotations to eliminate the interference of false logical forms and avoid the parser from falling into a local optimal solution.(3)Set up multiple sets of comparative experiments.Verify the effectiveness of weakly supervised semantic parsing methods and improve system performance by combining active learning strategies.Compare the impact of different selection strategies and labeling schemes.(4)Construct a question and answer system for complex and complex questions.Based on the question answering method proposed in this thesis,a question answering system based on knowledge graph,which aggregates complex questions,is designed and implemented.It has been tested to meet the various requirements of the system.