Research On Semantic Understanding And Answer Generation Method For Intelligent Question Answering System

The development of modern network technology has effectively promoted the in-depth research and wide application of artificial intelligence technology.As an important branch of artificial intelligence technology,the question answering(QA)system is used to capture the user question,and then,the accurately identifies answers are fed back to the user by performing insight on the semantics of the question.It has attracted great attention on both academia and industry recently.Furthermore,knowledge graph(KG)provides high-quality datasets,which accelerates the research and application of intelligent QA system based on KG.The system performs semantic understanding of questions,and then,finds answer entities of given questions through inferring and reasoning in KG.However,due to the problem of missing relation between those entities in KG,the performance of QA system is not optimistic enough,which limits the universal application of QA over KG.To tackle this problem,this dissertation designs a novel semantic understanding and answer generation method for the intelligent QA system.Specifically,the theoretical and methodological research on question semantic understanding,entity answer reasoning,and natural answer generation has been carried out.The main innovations are shown as follows:(1)Research on the question understanding method using semi-supervised multi-scale convolutional neural network(CNN).Aiming at the problem of question understanding of QA system,a new semi-supervised multi-scale framework with CNN is proposed,which aims to solve the problem of data labeling and feature extraction.Firstly,in view of the difficulty of obtaining label data,the two-view embedding algorithm is used to train and obtain label data from unlabeled data,which reduces the labor cost of label data acquisition.Secondly,considering the difficulty of single convolution kernel feature extraction in CNN,a multi-scale CNN algorithm is proposed,which uses multiple convolution kernels of different sizes for convolutional operations in the same convolutional layer.In this process,each size of the convolutional kernel can obtain a dimension of feature information.The feature information of all dimensions generated can be spliced and recombined.Convolutional layer can extract more feature information hidden in the data.Finally,the simulation results show that the proposed algorithm effectively solves the problem of difficulty in obtaining feature information.Moreover,it can more essentially describe the internal features of data and better capture semantic information.(2)Research on the entity answer reasoning method based on relation prediction.Aiming at the problem of entity answer reasoning of QA system,a method of entity answer reasoning based on relation prediction is proposed,which solves the problem of relation completion and entity answer generation.Firstly,aiming at the problem of missing relations in KG,an entity importance estimation method based on attention graph embedding is proposed,which can infer the initially hidden latent relations from the existing relevant triples,and finally realizes relation prediction in KG.Thanks to the completion of missing relationships in KG,entity answer reasoning based on the entity and the relation is performed to obtain the corresponding entity answer.Secondly,in order to further optimize the candidate entity answer,this dissertation uses a semi-supervised multi-scale CNN to process the question.It implements deep semantic analysis of the natural language question and matches the similarity of the question and the candidate answer,resulting in the most similar entity answer to the question.Finally,the results show that the proposed method further improves the performance of entity answer reasoning while realizing relation completion.In addition,both the relation prediction module and the entity answer optimization mechanism greatly improve the performance of the entity answer inference model on different datasets.(3)Research on the answer generation based on multi-level copy and question-aware loss.Aiming at the problem of natural answer generation of QA system,a natural answer generation method based on multi-level copy and question-aware loss is proposed,which solves the problem of generating meaningless natural answers and question-answer mismatch.Firstly,aiming at the problem of existing methods generating universal responses and incomplete responses,a decoder based on a multi-level copy and prediction mechanism is proposed.It can replicate related semantic units in both question and triplet simultaneously,and then predict commonly used words from the vocabulary.A syntactical natural sequence of answers is generated.Secondly,due to the mismatch between the generated natural answer sequence and the input question,a question-answer matching mechanism based on question-aware loss is designed.It enables the natural answer generation model to generate target answer sub-sequences that match the question word of the question,and finally generate a natural answer corresponding to the input question.Finally,the simulation results clearly show that the proposed method can generate a natural answer sequence that conforms to the grammar and context according to user requirements.Furthermore,it has laid a theoretical foundation for the complex multi-turn dialogues.