ROUGE

ROUGE-N compares n-gram overlap, while ROUGE-L uses a longest-common-subsequence relationship; other variants introduce additional matching choices. Implementations may report recall, precision and F-style scores with different tokenization or stemming. The family originated in summarization evaluation, where overlap with human references can indicate coverage of expected content. ROUGE differs from a semantic encoder-based metric because many common variants depend on shared words or sequences. It also differs from a factuality check: an output can reuse the right vocabulary while assigning an action to the wrong person or reversing a relationship.

The practitioner selects the variant and implementation, records preprocessing and reference handling and validates the score against representative summaries. References should cover the content priorities of the task. Useful artifacts include metric settings, system comparisons and reviewed examples of high-scoring errors or low-scoring valid paraphrases. Factual consistency and required-field coverage are checked separately. When a summary length changes, the effects on recall and precision are examined instead of interpreting one number without context. Comparisons keep the same source and reference collection.

Reference summaries are not unique, and lexical overlap can penalize valid paraphrases or reward copying. Recall-heavy settings can favor longer outputs, while a high score can hide factual errors and missing critical qualifiers. Scores depend on variant and preprocessing, so reports should state them explicitly. ROUGE is useful as a baseline when interpreted alongside factual and task-specific checks, rather than presented as a universal measure of summarization or language-model quality.