Research On Machine Reading Comprehension Method Based On Deeping Learning

Machine reading comprehension is one of the hot research topics in the field of natural language processing,and its significance is to enable machines to understand text semantics and have the ability to reason,extract text information and answer related questions.With the continuous development of deep learning technology and the release of large-scale machine reading comprehension datasets,a large number of excellent models have been continuously proposed and improved.However,the traditional word2 vec word embedding technology cannot solve the problems of polysemy and long-distance dependence of semantics,and the semantic information of the context with a long distance cannot be well encoded;the semantic information of the question and the document is not sufficiently interactive and fused,so that the model cannot find the part of the document that is helpful for answering the question.In response to these problems,an ensemble learning method is proposed,which uses BERT-wwm-ext as the text encoder and multiple heterogeneous deep learning models as the base model.The validity of the model is verified on Chinese Squad dataset.The main work is as follows:（1）Build base classification ensemble modelIn order to solve the problems in traditional natural language processing（NLP）,the encoding layer cannot obtain the information of processing context,the polysemy of a word and the feature of Chinese word segmentation are not considered.The BERT-wwm-ext model released by i FLYTEK Labs of Harbin Institute of Technology was selected as the pre-training model,and combined with the advantages of the classic model in the field of machine reading comprehension,a volume based on the attention mechanism and the dimension reduction of model parameters through convolution was proposed.It also includes a product network network,a lightweight convolutional network based on a self-attention mechanism,and a Bi-LSTM model based on a bidirectional long-term and short-term memory network supplemented by an attention mechanism.The convolution-based model improves the abstraction ability of the model,while the bidirectional LSTM has strong semantic extraction ability,and combining them can achieve complementary advantages.（2）Improved stacking ensemble learning methodAiming at the improvement of the classic model,an improved ensemble learning method is proposed to integrate three machine reading comprehension models to improve the effect of the ensemble model.The output layer of the base model is removed,and the output features of the base model are directly input into the meta-model for fusion learning.The experimental results show that the ensemble learning method can effectively improve the recognition effect of the model.