Contrastive sentence representation learning with adaptive false negative cancellation

Lingling Xu, Haoran Xie, Fu Lee Wang, Xiaohui Tao, Weiming Wang, Qing Li

Abstract: Highlights•The study presents novel methods with similarity metrics to identify false negatives in contrastive learning.•Adaptive false negative elimination and attraction methods are introduced, further boosting performance over SimCSE.•The proposed method demonstrates wide generalizability across various augmentation strategies and existing models.•The method outperforms external SimCSE in detecting false negatives and is more computationally efficient.