Research On Deep Agent Text Summary Generation Technology Based On Reinforcement Learning

With the continuous development of the Internet,cloud computing,5G technology and we-media industry,people generate and acquire more and more information every day,and most of the information exists in the form of text.Automatic text summarization technology can quickly help people get clear and readable information in massive texts.Aiming at the problem that traditional text abstracts based on deep learning cannot produce high-quality long text abstracts,a deep agent text abstracts model based on reinforcement learning is explored.The main tasks are as follows:(1)Faced with the problem that Recurrent Neural Networks tend to lose prior information when encoding long texts,a deep communication agent model for information sharing is proposed.This model divides the Long Short-Term Memory into different agents,each of which is responsible for encoding a portion of the source text.Each agent then broadcasts its own encoded results,allowing all agents to share global context information.After the final encoding is completed,the model integrates information from multiple agents through a contextual attention mechanism and passes it to the decoder to optimize the process of generating long text abstracts.(2)In order to make the topic words appear in the generated abstract as much as possible to improve its quality,a joint attention mechanism which can add topic information into the model is proposed.Firstly,word2 wector technology and Convolutional Neural Network are used to encode words and text segments respectively.Then,Euclidean distance between the two is calculated to extract topic words.Finally,the topic attention mechanism is used to combine the topic words embeddings in a linear way to form a joint attention mechanism with the contextual attention mechanism.By means of bias probability generation,the probability of the occurrence of the topic words in the final generated summary is improved.This model is tested on the CNN/Daily Mail and NEW YORK TIMES data sets.To solve the exposure bias problem,the network is trained end-to-end using a mixed training method with reinforcement learning,and the summaries are assessed using a recall rate-based ROUGE evaluation index.The experimental results show that the long text abstract generated by the trained model not only gets a higher score,but also has good generality and high readability,which verifies the progressiveness of the deep agent text abstract model proposed in this paper based on reinforcement learning.