Research On Machine Reading Comprehension Based On Deep Learning

Machine reading comprehension(MRC)requires the machine to answer the answers according to the given paragraphs and questions.The MRC task reflects the current level of development of NLP to a certain extent.In recent years,pre-trained language models such as BERT have made outstanding progress in many tasks,including machine reading comprehension.However,there are still problems such as the inconsistency between the pretraining process and the reading comprehension task format,which introduces errors,the inconsistency of model training objectives and evaluation indicators,and the model's lack of reasoning ability and reverse thinking ability.This thesis proposes a reading comprehension model that integrates multiple thinking based on the pre-trained language model.The model fully explores the relationship between paragraphs,questions,and answers,and redefines reading comprehension as a two-stage task that is generated and then selected.The answer generation model improves the encoder,decoder,and answer generation method based on the Sequence to Sequence model.The answer selection model uses comparative learning to improve the final answer selection method.The answer generation model is based on the Sequence to Sequence model to improve the encoder,decoder and answer generation methods.In order to verify the effectiveness of the model,this thesis designs corresponding multiple cognitive neural networks for the Chinese machine reading comprehension data set Du Reader.For the encoder part,firstly,the intensive reading module is proposed for the problem of introducing errors by concatenating paragraphs and questions when coding the machine reading comprehension task through the pre-trained language model.The intensive reading module can perform secondary encoding on the output vector of the pre-training language model according to the attention mechanism to obtain a paragraph vector that is strongly related to the question.Secondly,a reasoning module is proposed by simulating the human reasoning process.The question vector and the vector encoded by the intensive reading module are reasoned through multi-step calculations to obtain the final reasoning result.The encoder part infers from the forward and reverse directions to obtain the result vector and combines the result vector through a particular ratio to simulate the auxiliary effect of reverse thinking on forward thinking in human reading comprehension.The decoder part adopts a Transformer-like decoder and introduces mutual attention calculation with the problem vector.The model maintains a high degree of attention to the problem at every step in the decoding process.The character generation method uses a pointer generation network to restore the details in the paragraph to the greatest extent.The answer selection model uses comparative learning to compare the answer vector generated by the answer generation model and the "standard answer" vector generated by the prompt template according to the actual evaluation index order,so that the model can learn what kind of answer can get a higher score after logical thinking Make the choice of the answer.In the end,the model's ROUGE-L and BLEU-4 indicators on the DuReader data set reached54.71 and 40.46,respectively.Compared with the baseline model,the ROUGE-L score increased by 14.2%,and the BLEU-4 score increased by 4.1% and passed the ablation experiment proved the validity of each part of the model.