Which statement about BLEU is correct?

• A. Perplexity is derived from cross-entropy.
• B. BLEU scores are not based on references.
• C. BLEU score measures n-gram overlap with references.
• D. BLEU scores reflect semantic adequacy perfectly.

Answer: (C)

BLEU evaluates translation quality by comparing a candidate translation to reference translations through n-gram overlap. It looks at how many 1-, 2-, 3-, and 4-grams in the candidate appear in the references (with clipping to avoid inflating counts from repetition) and combines these into a score, typically with a brevity penalty to discourage outputs that are too short. This design makes BLEU fundamentally about surface-level overlap with references, not about measuring perplexity or cross-entropy, and not about guaranteed semantic fidelity. It does rely on reference texts for comparison, and it won’t perfectly capture whether the translation preserves meaning or style in every case.