Research On Abstract Generation Technology For News Text

The rapid development of the Internet and social media has led to the explosive growth of online news text data.How to obtain the required information quickly and efficiently has become an urgent problem to be solved.Automatic text summarization technology can compress long texts into short texts that are concise and coherent without losing the original meaning,thereby speeding up people's access to information.Generative text summarization models based on the Seq2 Seq model have achieved good results,but most of these models are trained in a teacher-forced way,which can lead to exposure bias problems.To this end,this paper explores two different solutions:(1)With the help of contrastive learning technology,by comparing the target sequence with its corresponding positive and negative samples in the feature space to learn a more reasonable prediction representation distribution;(2)With the help of adversarial reinforcement learning technology,when training a summary model,the selection of the next word is aimed at maximizing the reward of the future summary sequence,rather than directly improving the probability of the word through maximum likelihood estimation.The main contents of these two research works are:(1)Research on generative text summarization model based on contrastive learning.A pointer generator network PGN-CL based on adversarial perturbation contrastive learning is proposed to model the text summary generation process.The model takes the pointer generator network PGN as the basic architecture,and designs a new contrastive learning method to construct positive and negative samples,so that the model can be fully exposed to various correct and incorrect outputs during the training process to solve the exposure bias problem.Compared with randomly selected positive and negative samples,the positive and negative samples generated by the method in this paper are more difficult to distinguish,which allows the model to better learn the distinguishing features of positive and negative samples in the feature space and obtain a more accurate summary representation.The experimental results show that the PGN-CL model outperforms the baseline model on the ROUGE evaluation index,and can generate more accurate,coherent,and comprehensive summaries,which proves the effectiveness of introducing adversarial perturbation contrastive learning to improve the quality of summaries.(2)Research on generative text summarization model based on adversarial reinforcement learning.A pre-trained language model PLM-RLGAN based on adversarial reinforcement learning policy gradients is proposed to model the text summary generation process.The model models the generator and discriminator as agents and environments in reinforcement learning,respectively,through reinforcement learning policy gradients Implement generator optimizations for high reward summaries.During the model training process,the reward is used to decide which word to choose can make the quality of the summary prediction sequence the best,instead of maximizing the joint probability distribution of the words in the summary sentence through maximum likelihood estimation,so as to effectively avoid exposure bias question.The training method is applied to the pre-trained language model and the more advanced text summarization model,and the effectiveness of the adversarial reinforcement training on the improvement of the text summarization model is proved.