When a machine writes a sentence, how do we score it against a human REFERENCE? BLEU counts how many of the candidate’s word-chains (n-grams) appear in the reference — with a clip so you can’t win by repeating ‘the the the’, and a brevity penalty so you can’t win by saying almost nothing. Slide the candidate from garbage to a match.
BLEU scores a candidate against a reference by MODIFIED n-gram precision — matched n-grams clipped to their max count in the reference (so repetition can’t inflate it) — multiplied by a brevity penalty BP = min(1, e^(1−r/c)) that punishes candidates shorter than the reference. A perfect copy scores 1; a spammy or too-short candidate scores low. It is the long-standing (if imperfect) yardstick for machine translation and generation. A fail-loud self-check throws unless an identical candidate scores 1 and a repetitive one scores below 0.5. ◆ real NLP metric, node-verified.
Shown with unigrams for clarity (full BLEU geometric-means 1…4-grams); the clip + brevity penalty are exact. BLEU rewards surface OVERLAP with one reference — a good paraphrase using different words scores low, a known limitation.