Research On Accuracy And Real Time Response Of Question Answering System Based On Neural Network

Question answering system generally refers to the comprehensive agents that understand,analyze and finally give answers to the problems described by human natural language according to the existing resources.Question answering system has been widely used in many fields.Question answering system can be divided into question answering system without external knowledge and question answering system with external knowledge according to the different data used.Question answering system without external knowledge is oriented to more general fields,in which there is a large amount of question and answer data,which is often answered by retrieval.Question answering system with external knowledge is mainly applied in special professional fields.It needs to use existing knowledge to build a question answering system,which usually needs to be built based on the specific characteristics of knowledge.The accuracy and timeliness of reply are important indexes to evaluate the quality of a Q&A system.According to the characteristics of question answering system,this paper studies from the following three aspects:Research on the reply accuracy of retrieval question answering system without external knowledge.Realized the text fast retrieval in the context of big data.The candidate set of similar text is obtained by constructing inverted index and BM25 algorithm.By using the semantic similarity calculation model based on BERT,the candidate set is reordered and the best retrieval result is obtained.The experimental results show that the introduction of Bert model can effectively improve the accuracy of reply of the question answering system.Research on response accuracy of database knowledge oriented question answering System.Achieve a high level of understanding of database knowledge.The problem and the database header are used as the input of XLNET,and the header text information is introduced in the input side to strengthen the model's understanding of the database knowledge characteristics and the relationship between the problem and the knowledge.It realizes the efficient parsing of semantic expression information.Based on the characteristics of SQL statement,the single task is divided into six interrelated sub-tasks to reduce the difficulty of model parsing.Experimental results show that the model effectively improves the accuracy of recovery,far exceeding the baseline model.Real-time study of question-answering system.In order to improve the real-time performance of the Q&A system,the pre-training model used in the Q&A system is distilled.A migration distillation method is implemented,which distills the sentence vectors generated in the pre-training stage,and fine-tuning the distillation model obtained in the downstream task.The experimental results show that,on multiple tasks,the distillation model can improve the real-time performance of the model and reduce the memory consumption while retaining the performance of the original model,which effectively improves the real-time performance of the question answering system.