Research And Implementation Of News Article Summarization Technology Based On Deep Learning

With the rapid development of artificial intelligence,text summarization can liberate people from complex text information,but limited by various conditions,text summarization still has many problems.This paper focuses on the existing research difficulties and key requirements of text summarization.Under the framework of extractive-abstractive summarization,in view of the fact that redundant information affects the quality of summaries in the case of long text,and the two problems that the mainstream summary methods are easy to generate factual errors that affect the quality of summaries,a joint summarization model based on deep reinforcement learning is proposed,and the development and implementation of the summary system based on this model is completed.The main contents of this paper are as follows:(1)Construction of joint summarization model.Based on the encoder-decoder framework,this paper constructs a joint summarization model,which is an organic combination of extractive model and abstractive model.In order to solve the problem that the joint summarization model can not be trained effectively,the self-critical strategy gradient method of deep reinforcement learning is introduced in the process of joint training.In order to solve the problem of exposure bias caused by the inconsistency between training and testing standards,the ROUGE score is introduced into the enhanced reward to improve the accuracy of the model.(2)A de-redundant extractive summarization model is proposed.In order to solve the problem of low accuracy of one-time extraction in traditional methods,the one-time extraction process is modeled as a periodic extraction process based on Markov decision process to improve the accuracy of the model.In order to solve the problem of insufficient sentence vector information,a sentence vector encoder with the combination of local and global information is used to make the sentence vector pay attention to the global information and improve the representation ability of the model to the sentence information.In addition,a de-redundancy module based on multi-head attention mechanism is proposed to make the extraction model pay more attention to the important content of non-repetition and reduce the impact of redundant information on the quality of the summary in the case of long text.(3)A factual consistency enhancement method based on fact triple is proposed.In order to solve the problem that it is easy to generate factual errors due to the lack of additional information guidance in the abstractive model,the factual triples are extracted from the original text by Open IE tool to provide factual guidance for the generative model and reduce the probability of generating false facts.In order to solve the problem of insufficient representation ability of factual triples,the graph convolution network is used to capture the semantic and hierarchical relationship between factual triples,so as to improve the representation ability of factual triples.In view of the fact that only a small part of the factual triple information contributes greatly to the generated results,but there is a lack of related research at present,a factual triplet importance evaluation network is proposed to distinguish the contribution of different triples to the generated results.in view of the factual errors that may be caused by the introduction of loanwords in traditional decoding methods,the decoding method is improved by combining beam search and pointer network to generate the original words as much as possible.(4)text summarization system for news field.According to the research results of text summarization in this paper,a text summarization system for news field is designed and implemented.