ROUGE Score
ROUGE score is an automated metric, most common in summarization testing, that measures how much of a reference text's content (words, phrases, or word sequences) is captured in a generated summary — emphasizing recall of the source's key content over exact wording.
ROUGE comes in several variants (ROUGE-N for n-gram overlap, ROUGE-L for longest common subsequence) that trade off differently between rewarding exact phrase matches and rewarding capturing the right content in a different order — picking the wrong variant for a task is a common, quiet source of misleading scores.
Like BLEU, ROUGE has no real understanding of meaning — a summary that captures the right idea in different words scores lower than one that copies phrases verbatim from the source, even if the paraphrased version is actually the better summary a human would prefer.
It's best used as a fast regression signal during development (did this prompt or model change measurably shrink content overlap with reference summaries?) rather than as the final word on summary quality, which still needs human or LLM-judge review for coherence and accuracy.
