Research On Generation Of Reading Comprehension Subjects Based On Deep Learning

Proficiency in English reading comprehension questions is to assess learners knowledge of effective and necessary tools,however design related education significance of the topic is still human cost,thanks to the rise of deep learning,the problem of how to generate the appropriate and beguiling multiple-choice options,in the field of natural language processing is widely studied.However some difficulties still exist in the problem based on deep learning and interference item generation,currently contains answers to the questions generated,failed to make full use of the global information,and used for manual annotation of QA(Question Answer)is still very expensive,in addition to that,for the interference of multiple choice items generated,failed to fully use the interaction between the original text,issues,and the correct answer,leading to generate interference item confusing enough.Therefore,this paper solves the above problems through two models:1)This paper proposes a problem generation model based on answer extraction.In order to solve the problem that the generated questions contain the answer,this paper uses the method of answer substitution and prediction using semantic matching to solve the problem.The experimental results show that our model captures the target of the problem well,and the generated questions rarely contain the target answer.At the same time,we use the pre-training model UniLM(Unified pre-trained Language Model)to obtain more contextual information features and improve the accuracy and fluency of the model.Finally,we combined the answer extraction module composed of Bert(Bidirectional Encoder Representations for Transformers)+BILSTM(Bi-directional Long Short-Term Memory)+CRF(Conditional Random Fields)and used the model to automatically generate QA to train the MRC system.The results show that this QA pair is an effective training set that can help improve the performance of model training and reduce the burden of manual annotation.2)This paper proposes to use multiple choice model generation based on distractor attention mechanism.In order to solve the problem of failed to fully using the correct answer,and the original relationship between first BERT in the process of the training model is adopted to encode the original problem and,he can get more contextual information,more advantage in reasoning,generated after machine measurement shows that the interference of richness and smooth degree has a certain degree of ascension.Secondly by the interference of double change information extraction,makes the prediction of decoding phase disturbance item sequence,stress and problems associated with the original information,and do not use the answer related information,so as to generate more disruptive problem,through artificial three dimensions test and measurement of the machine on public data sets show that our model generated interference item more coherent and more education significance.