Research On Extractive Reading Comprehension Based On MGRU

With the emergence of recurrent neural networks,natural language processing tasks such as named entity recognition,automatic question and answer,machine translation,text implication recognition have achieved good results.Text implication recognition and reading comprehension belong to the category of natural language understanding.Reading comprehension,as one of the most difficult tasks in natural language processing,can be applied to automatic question answering and intelligent search.With the emergence of reading comprehension corpus SQuAD and TriviaQA web,it is possible to study reading comprehension in neural networks.LSTM,as the preferred model for text feature extraction,promotes the development of natural language tasks.For the task of text implication recognition,the focus is on the matching of subevents in each text.In this paper,word-by-word matching attention model and mLSTM model are studied,and the shortcomings of the two models are pointed out.This paper improves the current optimal mLSTM model and proposes a GRU-based adjacent word information and word-by-word matching model mGRU.The main idea of mGRU is to set a text or sentence as a premise,another text or sentence as a hypothesis,first coding the premise,then coding the hypothesis,and introducing the attention of all the words of the premise while coding the hypothesis.Inspired by the idea of n-gram,the hypothetical adjacent words hide state information and match preconditions word by word,which has achieved good results in identifying SNLI corpus in Stanford Text Implication Recognition,and is currently a better level of feature extractor based on cyclic network.For single paragraph extraction reading comprehension task,the basic research is done from text implication recognition.Considering that text implication recognition is the basic task of natural language understanding,the field of text-text relationship recognition can be analogously used in the encoding stage of extractive reading comprehension of a single paragraph,taking the problem as the premise,the paragraph as the hypothesis,and improving and optimizing the mGRU model.There are two improvements.One is to combine the boundary model of the pointer network with the boundary model as the prediction layer of the answer.The other is to match the output of the word-by-word matching model with two-headed self-attention.The improved model achieves good results in extractive reading comprehension of single paragraph.For the multi-paragraph decisive reading comprehension task,because SQuAD datasets are all from Wikipedia articles,the range is very small and the answer of short text is very short.True use in reading comprehension tasks is not credible,and the effect on more complex data sets is unknown.So the more complex reading comprehension tasks are studied.The single-paragraph optimal model is applied to the complex multiparagraph data set TriviaQA web.A paragraph selection method and a noise processing method are proposed for complex data sets.Combining these two methods,the model achieves good results in multi-paragraph extraction reading comprehension tasks.