A Span-Extraction Machine Reading Comprehension For Open-Domain Question Answering

As the Natural Language Processing develop,Open-domain Question Answering has become an increasingly important research topic.It is based on the user's natural language questions,which can accurately find the answers from a large number of candidate texts,and the Machine Reading Comprehension technology for Open-domain Question Answering System has become a hot research direction.Machines can understand human text language more accurately,so as to construct the Open-domain Question Answering Systems,the Automated Reading Comprehension Evaluation Systems and the Electronic Question Answering Systems.This is of great significance to the research in the field of Learning Analysis,to assist teachers in answering questions,marking papers,student intervention,and to help students predict their performance and learn more efficiently.In view of the slow training and prediction speed and low prediction accuracy of most current MRC models,how to build faster and more accurate models,especially the MRC model based on deep learning,is the key task of current research.Therefore,this paper proposes a network structure of QA-Reader based on pre-trained language model,deep neural network and multi-layer attention mechanism,and designs a Chinese Open-domain reading comprehension Q&A system based on free text.Specifically,the main research work of this paper is as follows.First,build a model of MRC.In this paper,a span-extraction MRC model QA-Reader is designed and implemented.Firstly,we use RoBERTa-wwm-ext to obtain the word embedding representation of the text and the question.Then we use deep separable convolution and multi-head self-attention mechanism to encode.Then the bidirectional attention of the text and the problem and the self-attention of the text are calculated to get the final semantic representation.Finally,the answer is predicted.For the unanswerable questions,the model also calculates the probability of unanswerable questions.Finally,the model is tested on two Chinese span-extraction MRC datasets,and the results show that the proposed model has higher accuracy and faster training speed through the comparison of various experimental schemes.Second,design Chinese Open-domain QA System.In this paper,the MRC model is applied to the Open-domain QA System.Not like the previous system,the system in this paper can complete two tasks of reading comprehension and open QA.The reading comprehension task implements basic reading comprehension based the QA-reader model;Open QA task implements open field QA based on free text,and the system uses Chinese Wikipedia as the only answer source.In this system,the MRC model,as a supporting key functional module,can complete the basic MRC task.