Research On News Article Summarization Based On Reinforcement Learning

News has always been an important way for people to retrieve information.Espe-cially with the popularization of mobile smart terminals,many emerging media such as Weibo and WeChat have gradually replaced traditional paper media.In order to attract people's attention,these new media often exaggerate their titles,but they may have noth-ing to do with the content.The current fast-paced lifestyle determines that reading is often fragmented,and people urgently need a way to understand the key content of news in a short period of time.Automatic text summarization can compress and summarize news,extract key content,filter redundant information,and improve people's reading efficiency.This thesis combines reinforcement learning methods to improve the text summarization model.The main research contents are as follows:(1)Construction and improvement of generative summary model.Adopt Encoder-Decoder framework,use pre-trained language model to extract semantic features,and use internal attention mechanism to memorize historical key information to solve the problem of long-distance dependence;increase the pointer mechanism to solve the problem of unregistered words.(2)Based on the self-critical strategy gradient method in reinforcement learning,the quality of the abstract generated by the model is improved.Mainly improve the perfor-mance of two aspects:one is the performance on the ROUGE score;the other is the semantic coherence of the abstract.Among them,This thesis proposes a semantic coher-ence evaluation network based on a pre-training model to score the semantic coherence of the generated abstract.Finally,experiments were conducted on the CNN news data set.The experimental results confirmed that the self-critical strategy gradient method can ef-fectively improve the performance of the generative model on the ROUGE indicator and semantic coherence indicator.(3)To solve the problem of insufficient ability of the generative summarization model to deal with long texts,we propose a generative summarization enhancement method based on entity features and actor critic algorithm.First use the extraction model to ex-tract important sentences,and then use the generative model for summary generation.In terms of extractive models,it is proposed to introduce co-referential entity information in the text to assist model extraction to enhance the accuracy of extracting sentences.At the same time,based on the actor critic algorithm in reinforcement learning,the two models are formed into an end-to-end system for training to maximize the advantages of extraction and generation.The experimental results also verify the effectiveness of the generative summary enhancement method based on the actor critic algorithm.(4)Designed and implemented a news article summary generation system.Based on the summary model proposed in this thesis,combined with Web development technology to realize a news article summary system,provide users with news summary services,and display the final results to users on the web.