Machine Reading Comprehension Model Design Based On Specific Dataset

Natural language processing,as a promising field of current artificial intelligence,mainly explores the interaction between natural language and computers.Making machines have the same reading comprehension ability as human beings has attracted the attention of many researchers.If machines are capable of a fairly high level of reading comprehension,then many applications will show true intelligence and serve humans better.There are many forms of machine reading comprehension,and fragment extraction type is a hot area of research.Various high-level models based on deep learning have been created.On the SQuAD dataset,for a given question,the answer is limited to a fragment of the text,and the basic architecture of existing models can be summarized as four parts: embedding layer,encoder layer,attention layer and output layer.The encoder layer mainly uses recurrent neural network to encode the original text and question,so that the vector representation of each word after encoding contains semantic information of the context.This paper attempts to encode text and problems at the encoder layer using a convolutional neural network.In the experiment analysis part,based on the SQuAD dataset,a baseline model containing the encoder layer of recurrent neural network is first implemented,in which a sliding window is used in the output layer to prevent the model from outputting empty when forecasting.Secondly,based on the statistical analysis of the training dataset,the improved model implements an encoder layer based on multi-layer convolution to replace the original encoder layer of the baseline model.Through experimental analysis,the performance of the improved model on the given evaluation metrics is slightly lower than the baseline model,but it is dominant in the number of parameters and the iteration speed,and the model design is reasonable.