Research On Chinese Text Summarization Based On Deep Learning

The continuous development of Internet technology has enabled people to access all kinds of text information in various convenient ways.To improve the efficiency of information acquisition,the technique of text summarization has been studied intensively and extensively in recent years.Currently,most of the text summarization techniques are implemented by deep learning algorithms.The main research of this paper is Chinese text summarization based on deep learning.This topic proposes a solution to the problem that feature suppression affects the quality of summary generation.The model is also optimized using text structure features and the copy mechanism of the predicted labels.Feature suppression refers to the inability of the model to distinguish between textual similarity and semantic similarity in text-based tasks.In response to the problem that features suppression affects the quality of summary generation,This paper proposes a contrast learning method using random word order reference summaries as negative samples,exposing the model to text that is similar in structure but not in semantics,and building a new loss function so that the model can distinguish between text similarity and semantic similarity.In the structure of the text summarization model,this paper is based on the Seq2 seq architecture.A pre-trained language model,Ro BERTa Chinese Version,is used on the encoding side of the model;In terms of the structure of the text summarization model,the model structure in this paper is based on the Seq2 seq architecture.A pre-trained language model Ro BERTa Chinese Version is used on the encoding side of the model;Compression of data using residual inflation convolutional neural network;Copy mechanism by pointer generation network.Based on the above scheme,this paper designs and implements a residual dilated convolutional summary generation model based on contrast learning(RGCCL).The experimental results on the LCSTS dataset show that the ROUGE score of RGCCL is significantly improved compared with the baseline model,and the generated summary text can more accurately express the main idea of the original text,which proves that the method proposed in this paper solves the problem that feature suppression affects the quality of summary generation.Based on the RGCCL model,two new model optimization schemes are proposed in this paper: one is to use text structure features to assist summary generation,which improves the quality of w our generated summaries;the other is to use the predicted label copy mechanism instead of the Pointer Generation Network copy mechanism,which shortens the RGCCL model training time.The comparison experimental results show that the RGCCL model applying these two optimization schemes both improves the quality of summary generation and shortens the training time.Compared with the baseline summary model,the text summarization model implemented in this paper generates higher-quality summaries and has a shorter training time.By performing text summarization on recent news texts,the generated summary text results can accurately express the main idea of the original text,which proves that the text summarization model implemented in this topic has practical application value.