Research On The Effect Of Feature Representation And Model Ensemble In Machine Reading Comprehension

Machine reading comprehension in the field of artificial intelligence(AI)is one of the core tasks in Natural Language Processing(NLP).The aim of the research is to make the machine read the original text and correctly answer the questions related to the content of the original text.Improving the accuracy of the machine reading comprehension can increase the accuracy of machine translation and information retrieval,promote the research of Question Answer(QA)and other NLP tasks.At the same time,the representation of text features also has a certain influence on the accuracy of the machine reading comprehension.Therefore,the research of text feature representation and machine reading accuracy improvement has very practical application value.This thesis describes the process of text feature processing,and improves the expression of text features in the Mahmoud Nabil paper of 2016.For the words with "'" in English,the unreasonable way to treat the words as two words is changed to restore the complete abbreviation.At the same time,the filling word is used to remove the high and low frequency words.The way of occupying is replaced by another word.The original filling words are used only for filling.After the text features are quantized,the high and low frequency words are completely removed without changing the original text feature order.In the stage of model ensemble,the grammatical and semantic information of the text is characterized with the trained word vector or the position word vector,and the model ensemble is built on the basis of the existing classical neural network model.Finally,11 sub-models with an accuracy more than 65% were selected to improve the accuracy of machine reading comprehension by soft voting and hard voting.Finally,the results of the performance improvement of the machine reading comprehension task are compared and analyzed by using the sub-models of different structures on the training data to carry out the model ensemble.The results show that after the high frequency words are completely removed,the prediction effect of the model has a certain improvement.With the ensemble of multiple sub-models with relatively large structural differences,the ensemble result is obviously better than that of a single sub-model.