Research On Key Issues Of Text Summarization Based On Automatic Machine Learning

The advent of the 5G era has accelerated the vigorous development of the Internet and various smart terminals,greatly satisfying people's network needs.However,as more and more text data are generated,it is difficult for people to obtain the key information efficiently.This will not only affect people's reading experience,but also greatly increase people's loss of time and energy.Therefore,it is of great significance to study an efficient automatic text summarization technology.Aiming at the dialogue text,a special kind of text data,this thesis proposes a dialogue summary generation method based on the Neural Architecture Search(NAS)technology in the field of automatic machine learning.The main work is as follows:(1)Aiming at the problem of discrete topics in dialogue texts,this thesis proposes a topic division method for dialogue texts based on self-supervised learning and NAS technology.In order to divide the dialogue text into several paragraphs according to the topic information,a self-supervised dialogue text segmentation model is proposed,and a novel Recurrent Neural Network(RNN)structure is designed based on the NAS method of reinforcement learning to improve the ability of capturing the semantic information of the dialogue text for the segmentation model.Finally,a topic paragraph clustering algorithm is implemented for the generated discrete paragraphs,which can cluster paragraphs describing the same topic content together.Through experiments,it is shown that the method proposed in this thesis can effectively divide the topic information in the dialogue text.(2)In order to generate dialogue summaries based on divided topic paragraphs,this thesis proposes a dialogue text summarization generation model combining Transformer and Generative Adversarial Network(GAN),and designs the search space of generative and discriminative networks.The search method realizes the joint search process of the whole network,which is based on a differentiable NAS method.The effectiveness of the proposed method for generating dialogue summary is demonstrated through experimental comparisons on public dialogue datasets.(3)In view of the customer service system's demand for ticket summary generation,this thesis designs a customer service ticket summary system based on the proposed dialogue text summary generation method,and implements the related functions of ticket summary generation.The system is designed to help customer service staff efficiently record ticket summaries,thereby improving customer service staff's work efficiency.