Research And Implementation Of Selective Machine Reading Comprehension

With the development of deep learning technology in recent years,the field of natural language processing has also made rapid progress.Among them,machine reading comprehension,as an emerging field,has become more and more mature in these years,and has begun to appear in practical applications in life,including intelligent customer service,large Scale text processing,etc.Machine reading comprehension refers to the ability of computers to extract key information from written texts,understand the structure of texts,and allow computers to understand the content of texts just like humans.At present,the mainstream method of machine reading comprehension is obtaining the correct answer by letting the neural network learn the reading comprehension data set.Reading comprehension can be divided into fill-in-the-blank reading comprehension,extractive reading comprehension,and selective reading comprehension.Among them,selective reading comprehension is currently the main field of reading comprehension.The content of the options does not completely correspond to the original text.There are many reasoning types,so the difficulty will increase accordingly.The model is required not only to understand and learn the original content,but also to have a certain degree of logical reasoning.Starting from this,our article analyzes and experiments on the mainstream large-scale selective reading comprehension data set RACE,uses neural network models to learn to achieve machine selective reading comprehension,and studies a variety of strategies to help improve the test accuracy.In view of the current problems in selective reading comprehension,our article conducts research from the following two aspects.First,from the design perspective of the model itself.Based on the pytorch deep learning framework,our article improves the current mainstream pre-training models Bert and XLNet to achieve selective reading comprehension tasks.Second,from the perspective of strategic research.In view of the characteristics of the RACE data set and the behavioral performance of mining human reading comprehension,this paper conducts the strategy research and implementation of the external knowledge base in the introduction of external knowledge,and conducts the strategy research of the highlight embedding of keywords in the improvement of the model structure.Research and implementation of difficult sample re-learning strategy in the aspect of RACE data sample learning improvement.Applying the three innovation strategies to the Bert and XLNet network models has improved the accuracy of the test set to varying degrees.Through testing on the RACE data test set and the evaluation indicators is accuracy rates.Applying the three innovation strategies to the Bert and XLNet models respectively,the test accuracy has been improved to a certain extent,and the effectiveness of the three innovation strategies is verified.Performing mixed tests on the three innovation strategies to verify that the three innovation strategies will not have a negative impact due to the mixed use.Any mixing will improve the test accuracy.The experimental results show that the three innovation strategies are carried out together in a mixed experiment,and the accuracy of the basic model based on Bert and XLNet has been increased by 3.3 and 3.4 percentage points respectively,which verifies the effectiveness of the strategies proposed in this article and can be applied to reading comprehension.In the corresponding application.