Research On Deep Learning-Based Machine Reading Comprehension Model

Machine reading comprehension refers to letting a computer read text like a human,extract text information and answer questions.With the rapid development of deep learning technology,the deep learning-based reading comprehension model has become a current research hotspot.However most of the existing machine reading comprehension models have the following shortcomings:(1)Traditional pre-trained word embedding technology can't solve the phenomenon of polysemy.(2)The BiLSTM is used for encoding.Although this encoding method has a good effect for short-distance dependency of text,it still can't effectively deal with the long-distance dependency,and the computing speed is slow due to the complexity of the structure.(3)There is not enough interaction and fusion between the semantic information of the question and the article,which makes the model can not find the helpful part to answer the question from the article itself.(4)The answer to the question depends entirely on the answer interval predicted by the model,which is a too single way to predict the answer.In view of the above problems,the research constructs two machine reading comprehension models,BiGSANet and RoBTANet,and conducts comparative experiments on the DuReader dataset.The major research tasks are as follows:First,the BiGSANet mode was developed.In modeling process,the BiGRU is used as a text encoder in BiGSANet,and after bi-directional attention model is used for semantic interaction,the semantic fusion layer is added,so that the model can better integrate the semantic information and output answer range of questions and articles.Finally,an answer selection algorithm is designed,according to different factors to verify the quality of each candidate answer,to choose the best answer.In addition,an auxiliary task is also designed to promote model training.Second,the RoBTANet model was developed.In the modeling process,RoBERTa coding in RoBTANet and multi-head attention mechanism in semantic interaction layer are used to interact with the question and the context information.Then,the Transformer coding block is used to make the semantic information of the question and the context be fully fused,and the possible answer interval is output,and the fully fused coding matrix is input into the auxiliary task.In addition,the second verification is added to the answer the selection of the answer to help the model of choice the answer to the answer selection algorithm to help the model choose the better answer.Third,the validate the effects of above models,two experiments on the large-scale reading comprehension DuReader are conducted.The results show that the coding performance and computational speed of using RoBERTa is much better than that of BiGRU.The Rouge-L and BLEU-4 of RoBTANet model are 59.35 and 56.22 respectively on DuReader,which are significantly better than most existing machine reading comprehension models.