A training technique where negative examples (dissimilar samples) come from other items in the same training batch, helping the model learn to distinguish between similar and dissimilar texts.