Research On Constructing Automatic Question Answering System Based On Convolution Neural Network Of Semantic Matching

With the rapid development of information technology,users' access to network resources are constantly changing.From initial yellow pages to traditional search engines,and then to current intelligent question-answering robots directly providing answers,such changes are made thanks to the significant progress in natural language understanding(NLU)and information extraction(IE).Users are increasingly inclined to use a simpler way to obtain information,which also requires computers to have stronger language-understanding ability.In recent years,more and more technology companies and research institutions have started to develop intelligent question-answering robots,such as Siri,Apple's intelligent voice assistant and Xiaoice launched by the Microsoft Search Technology Center,and the automatic question-answering module is a critical module of intelligent question-answering robots.Therefore,the automatic question-answering field has become popular in artificial intelligence researches.Besides,various new machine-learning methods applied in this field have also raised the intelligence level of question-answering systems.The purpose of this research is to build an automatic question-answering system for frequently asked questions.This thesis mainly includes the construction of open semantic-matching corpus,the design of semantic-matching algorithm,and the construction of an automatic question-answering system.In view of the fact that there is no open Chinese corpus in the field of semantic matching,and the most widely used MSRP corpus has only limited amount of data,this project constructed an open semantic-matching corpus that can be used in Chinese question-matching.Based on the corpus constructed,the project made some research on the algorithm for short-sentence semantic matching.By expressing questions through word embedding,this project compared the merits and demerits of the traditional similarity-based algorithm,the algorithm based on convolution neural network and the convolutional neural network model based on attention mechanism.Then a choice was made.The improved attention-based convolutional neural network algorithm has the advantage of extracting high-level abstract language features and automatic selection of some effective underlying features,which is superior to other methods' performance.Based on the above semantic matching model and the traditional techniques of information retrieval and semantic analysis,this project built an automatic question-answering system for the frequently asked questions of Alibaba's internal domain.The original experimental corpus of this project is mainly from Baidu Knows and Alibaba Cloud customer service.After being processed,this part of the corpus became a standard data set for model training.The results of the experiments show that the convolutional neural network model based on attention mechanism performed the best,with the F1 score reaching 78.3%.This project constructed an automatic question-answering system for e-commerce in the application stage.The docker service was used to deploy the system,and the accuracy of system return reached 84.7%.