A training technique that improves embeddings by comparing a text sample against multiple negative examples, helping the model learn to distinguish similar from dissimilar content.