Research On Machine Reading Comprehension Algorithm Based On Deep Representation Learning

Machine reading comprehension is a technique that uses algorithms to make a computer understand the passages and answer the relevant questions.The improvement of computer computing power and storage technology and the unique feature extraction ability of deep learning,the use of end-to-end technology to build machine reading comprehension models has become the main research method.The development of machine reading comprehension based on Chinese text is relatively slow due to the difficulty of constructing data sets.Baidu’s WebQA data set has filled this gap.The BiDAF model has excellent performance capabilities on the English SQuAD data set.This thesis attempts to transfer it to the WebQA data set to obtain the Chinese machine reading comprehension model BiDAF,and improve the coding layer and interaction layer of the existing model to improve the model’s WebQA The F1 value and EM value on the data set are the research goals.The main content of the study are as follows:This thesis first reproduces the classic BiDAF model.Apply it to the WebQA data set to get the baseline model of this thesis.Then,on the basis of BiDAF,this thesis improves the coding layer and interaction layer.For the coding layer of the model,in order to avoid the problem of boundary segmentation,this thesis proposes a machine reading comprehension algorithm model based on word hybrid embedding.In order to fully consider the entity attributes of words,this thesis integrates entity vectors to further strengthen the expression of word attributes,and add the pre-training language model to get the word vector representation.Through experimental analysis,the improvement of the coding layer improves the F1 value of the model.For the interaction layer of the model,this thesis proposes a reading comprehension algorithm model of gated dilated convolutional neural network based on attention interaction.CNN coding sequence is used,expansion convolution is used to expand the receptive field,and a gating mechanism is added to control the inflow of information.This coding method is recorded as GDCNN(Gated Dilated Convolutional Neural Net-work)coding.Use hierarchical attention and self-attention mechanisms to interact with passages and questions.The experimental results show that the semantic analysis ability of the improved model is improved.For the above two improvement methods,this thesis has done five sets of experiments.Through the comparison and analysis of the results on the verification set,it can be seen that the effect of the improved model has been greatly improved.Finally,based on the above two improved methods,this thesis proposes a hierarchical attention and self-attention gated dilated convolutional neural network HASA-GDCNN(Hierarchical Attention and Self Attention Gated Dilated Convolutional Neural Network)reading comprehension model.The performance of the model is verified on the data set WebQA,and compared with the public model,the F1 value of the HASA-GDCNN model is obtained.However,the EM value needs to be improved compared with the best results.At the end of this thesis,three comparative examples of experimental results are given,it shows that compared with the baseline model BiDAF,the results of the HASA-GDCNN model are more accurate.