Contrastive Learning for Improved Abstractive Sentence Summarization

Text summarizing aims to overview the material while maintaining essential details concisely. Lately, summary tasks have been included in the contrastive learning process in addition to visual representation. To optimize their similarities, contrastive learning summarizing algorithms currently concentrate on modelling and semantic similarity in all documents, targets, and data summaries. They do not, however, consider the impact that sentence semantics have on the text. In this study, we propose an RNN technique to develop the model for detecting the salient information. The sentence salience is based on the semantic similarity between the summary sentences of the original content, taking into account the distance of the loss of contrast in the form of soft weights functions. Consequently, our proposed model minimizes the semantic similarity given the textual summaries and possible noise and maximizes the semantic words between summaries and relevant information. Three popular datasets—multi news, CNN/Daily Mail, and Booksum datasets achieved the highest accuracy of 0.69% in the news datasets. The experiment outcomes demonstrate that the suggested strategy can considerably raise baseline performance and outperform other contrastive learning techniques.