Research On Machine Reading Comprehension Method Based On Deep Learning

Machine reading comprehension is an extremely challenging task in the field of natural language processing.with the core goal of enabling machines to understand human language and answer relevant questions.In recent years,with the continuous release of large-scale and high-quality reading comprehension datasets and the continuous improvement of deep learning representation capabilities,major breakthroughs have been made in machine reading comprehension research,which has important research value and broad application prospects.Machine reading comprehension can improve the efficiency and quality of human-computer interaction,and has been widely used in many fields such as intelligent question answering and information retrieval.Aiming at the extractive machine reading comprehension task,this paper is based on the QANet model,and improvements come from the aspects of feature extraction,text representation and semantic information interaction.The main research work is as follows:(1)A machine reading comprehension model based on rotational position encoding and half-step feedforward neural network is proposed.The QANet model is optimized by introducing rotational position encoding and half-step feedforward neural network.The rotational position encoding can better capture the semantic relationship and directionality information of text sequences,and solve the problem that some position information may be lost when absolute position encoding processes long sequence data.The half-step feedforward neural network can effectively capture the important features of the input data and alleviate the problem of gradient disappearance,thereby improving the training effect and stability of the model.The experimental results show that the optimization method based on rotational position encoding and half-step feedforward neural network can help improve the prediction performance of the model.(2)On the basis of introducing rotational position encoding and half-step feedforward neural network,a machine reading comprehension model based on RoBERTa embedding and multiple residual attention is proposed,and named Ro-QANet model.The introduction of the RoBERTa pre-trained language model effectively solves the problem of polysemy in traditional word embedding,and provides richer semantic information.Multiple residual attention helps to better integrate semantic interaction information,and solves the problems of insufficient interaction and weak relevance of semantic information between articles and questions.The experimental results show that,compared with the baseline model,the model improves the EM and F1 indicators by 4.7% and 3.8%,respectively,showing better prediction performance.(3)Based on the above research,a Web-based machine reading comprehension prototype system is designed.The system adopts the mainstream B/S architecture,combined with framework technologies such as Vue and Flask,and realizes functional modules such as data interaction,model calling,and answer prediction,providing users with efficient and convenient question-and-answer services.The experimental results show that all functional modules in the system can operate normally,which verifies the feasibility of the proposed model.