Research And Implementation Of Multi-document Machine Reading Comprehension Based On DuReader

In recent years,with the continuous development of deep learning technology and the continuous release of large-scale machine reading comprehension data sets,machine reading comprehension technology has made great progress.In 2018,Baidu released Du Reader,the first large-scale Chinese machine reading comprehension dataset so far,and the most representative large-scale Chinese machine reading comprehension dataset in the field of Chinese machine reading comprehension,and built two baseline models suitable for Du Reader dataset based on Bi DAF and MLSTM,which are two English machine reading comprehension models applied on the SQu AD dataset.After a year of development,Baidu increased the training set and test set of the Du Reader dataset to nearly 300000 and 120000 respectively,and Du Reader 2.0 dataset was born.The work in this thesis is only based on the Bi DAF baseline model,because the Bi DAF single model has achieved better performance than MLSTM on the SQu AD dataset.The main content of this thesis has three points:(1)In the word embedding layer,the pre-trained word vector is used to replace the randomly generated word vector of the baseline model to embed the model.(2)In the coding layer and the modeling layer,a multi-head attention mechanism is used to replace the Bi LSTM technology used in the baseline model for coding and modeling.The improved model is compared with the Du Reader baseline model on the Du Reader2.0 data set.The experimental results show that the improved model in this thesis is on the two evaluation index scores of ROUGE-L and BLEU-4,and the ROUGE-L score of the model in this thesis Compared with the Du Reader baseline model,it is improved by 3 percentage point,while the BLEU-4 score is slightly improved.At the same time,the training time of this model for an epoch is only about 2/5 of the Bi DAF baseline model,and only 1/4 of the MLSTM baseline model.It shows that the improved model in this thesis is a better model in terms of time cost and accuracy of predicting answers.(3)Use Python Web technology to design a Chinese machine reading comprehension system.The system can read and learn articles,and answer questions related to articles entered by users.