A contrastive learning technique that trains models to recognize when two slightly different versions of the same sentence are similar, improving semantic understanding.