Technical GlossaryNatural Language Processing

Hard Negative Mining

A training strategy that improves retrieval and matching quality by providing semantically confusing hard negatives rather than easy negatives.

Hard negative mining is one of the key determinants of representation quality in contrastive learning and retrieval training. Easy negatives often provide little learning value; the real gains come from examples that are semantically close yet incorrect. This makes training-data design nearly as important as the architecture itself. It can produce major quality improvements, especially in dense retrieval systems.