Dual-Level Contrastive Learning with Reverse Translation for Summarisation

Text summarisation aims to create brief and precise summaries from large amounts of textual information. Although recent advancements have improved summary brevity, significant room remains for enhancing ROUGE scores, especially with increase in beam width. We present a novel model, Dual Level Contrastive Learning with Reverse Translation (DC-RT), which utilizes contrastive learning with data augmentation, including reverse translation, to improve accuracy. Reverse translation serves as a key augmentation strategy by translating summaries to another language and back resulting in enriched training data and enhanced robustness of the model. Additionally, we introduce a new scoring function during training to further refine output quality. Exceptional ROUGE scores are achieved using this framework, ensuring concise and readable summaries even with increased beam width. Experiments on the CNN/DailyMail (achieving a47.82 ROUGE-1 score with a 0.017 variance) dataset demonstrates that there is a significant improvement in the precision and succinctness of summaries. Notably, our model surpasses the output of BRIO, the cutting edge model, setting a new benchmark in the field.