A Semantic Similarity Distance-Aware Contrastive Learning for Abstractive Summarization

Recently, contrastive learning has been extended from visual representation to summarization tasks. Abstractive summarization aims to generate a short description for a document while retaining significant information. 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 source document. In this paper, we propose a sentence-level semantic similarity distance-aware contrastive learning method (SSDCL), which integrates the semantic similarity distance between summaries and sentences of source documents 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 noise. We conducted extensive experiments on CNN/Daily Mail and XSum datasets to verify our model. The experimental results show that the proposed method achieved remarkable performance over the baseline and many advanced methods.