Effective contrastive losses in SRL depend on gradient dissipation, weight, and ratio components.
The author explores the effectiveness of contrastive SSL in Sentence Representation Learning (SRL) and identifies key components for optimization.