Research On Techniques Of Graph-Based Abstractive Text Summarization

With the exponential growth of Internet information,Internet texts become more diverse.Among them,the significant increase in the length and quantity of document-level texts has brought new challenges to the use of Internet information.As an effective information extraction technology,text summarization can extract important information from document-level texts accurately,which greatly improving the utilization of Internet text information.However,in addition to the larger length,the structure of the document-level texts is far more complicated than sentence-level texts.And the structural information is often closely related to the importance of the information in the text.How to use the structural information to recognize and represent the important information in the text effectively is a major challenge for the text summarization task.Another challenge of the text summarization is how to use the important information to generate summary accurately and concisely.In this thesis,we propose some methods to address both challenges.Specifically,the contributions of this thesis include the following three points:1.We propose a graph-based text representation method.The method can make use of the structural information of the original text to identify and represent important information of the text.In addition,by contrast with the self-attention model,this thesis also demonstrates the effective-ness of this approach theoretically;2.We introduce a salient sentence extractor into our abstractive summarization model.This module can select the salient sentences in the document effectively,which may improve the performance of summary generation;3.In this thesis,we propose a graph-based joint abstractive summarization model.In the training phase,the model regards extractive summarization and abstractive summarization as different tasks and adopts a training method similar to multitask learning.In the testing phase,the model can combine the advantages of extractive summarization and abstractive summarization to generate summary in a joint way.Finally,we conduct experiments on the CNN/Daily Mail dataset.Experimental results show that the graph-based joint abstractive summarization model can be effectively used for summary generation.Through a large number of quantitative experiments an d qualitative analysis,we verify the effectiveness of the model in various aspects.