Sentence salience contrastive learning for abstractive text summarization

Text summarization aims to generate a short summary for a document while preserving salient information. Recently, contrastive learning has been extended from visual representation to summarization tasks. At present, the methods of contrastive learning summarization focus on modeling the global semantics of source documents, targets and candidate summaries to maximize their similarities. However, they ignore the influence of sentence semantics in the document. In this paper, we propose a sentence-level salience contrastive learning method to help the model capture salient information and denoise. The model expresses the sentence salience according to the semantic similarity between the summaries and sentences of the source document, and integrates the similarity distance into the contrastive loss in the form of soft weights. Therefore, our model maximize the similarity between summaries and salient information, while minimizing the similarity between summaries and potential noise. We have verified our method in three widely used datasets, CNN/Daily Mail, XSum and PubMed. The experimental results show that the proposed method can significantly improve the baseline performance and achieve competitive performance in the existing contrastive learning methods.