| Automatic summarization technology aims to allow computers to summarize the information reflecting the main idea of the original text from the source text,and it is a supporting technology for the development of news headline generation and scientific literature,which can alleviate the problem of information overload.Among them,the single-document automatic summarization technology has attracted research by scholars at home and abroad with the development of large-scale data sets and deep learning,but there are also some problems that need to be solved in the existing abstract research.Therefore,this thesis addresses these issues by understanding semantic relationships based on strong sentence embeddings and coreferential relationships to improve the performance of downstream summarization tasks.(1)For the existing summarization technology,in order to reflect the style in the source text,there are many different representations of the same entity,resulting in vague summarization expression and poor readability.The second chapter proposes a co-referential digestion algorithm based on end2 end to preprocess text.This not only avoids confusion in the co-referential relationships of entities in the abstract,but also improves the relevance between sentences.The Bert Sum,Two Stage-Unified,Lo BART digest algorithms were combined to preprocess the source text on the news dataset CNN/Daily Mail,NYT,and the proceedings ar Xiv to verify the validity of the experiment.(2)Summarization model based on supervised learning often require a large number of high-quality annotated corpora,and the quality of the applied text summarization in different fields is greatly reduced once they are exchanged,and the semantic repetition of the extracted summarization is high.The third chapter proposes a summarization method of sentence representation graph model based on contrastive learning.The Sim CSE model based on contrastive learning captures deeper semantic information,and then converts the text into a graph structure to calculate the significance of sentence levels,and considers the influence of keywords on sentence-level significance from the characteristics of text structure.The validity of the experiment was verified on the news dataset CNN/Daily Mail,NYT and the proceedings set ar Xiv,and the recall rate of Rouge-1 reached 42.67%,44.58%,and39.58%,respectively,which was lower than the semantic repeatability of the traditional graph summary model.(3)For most abstractive summarization models based on a single document,it is difficult to reproduce the details of the text,and it is difficult to control the consistency of the summary with the facts of the source text.Chapter 4 proposes a sentence summarization method based on the combination of sentence level extractor and improved pointer generation network.In the first stage,the significance score of each sentence is obtained by using the graph-based model,and then the Read-Again mechanism is integrated to learn sentence embedding,imitating the habit of people summarizing the abstract by reading through the full text first and then summarizing the abstract,and when calculating the word-level attention distribution,it will be affected by the graph extractor to ensure the key content source,and the Coverage mechanism is further improved to alleviate repetitive content.Experimental results show that while ensuring the authenticity of abstracts,the recall rate of Rouge-1 in CNN/Daily Mail,NYT,ar Xiv datasets increased by0.30%,0.16%,and 0.69%,respectively. |
