NLP & Language · intermediate · concept 140 of 311 · Updated Nov 4, 2026

📐NLP Metrics (BLEU, ROUGE)

BLEU and ROUGE score generated text by how closely it matches human-written references: BLEU for translation, ROUGE for summaries. They are fast and repeatable, which suits regression testing. They also reward similar words rather than correct meaning, so a good answer in different words can score badly. For modern language models they are best paired with human or model-based judging.