Precision measures the proportion of items a system identifies as positive that are actually positive. It answers a practical question: when the system flags or selects something, how often is that decision relevant or correct? section 12.2.2
Use precision as part of a broader evaluation rather than as a standalone quality signal:
- Pair precision with recall: precision assesses the correctness of positive selections, while recall assesses how much relevant reference content or eligible material the system captures. section 12.2.2
- Use F1 when a balanced trade-off between precision and recall is needed. section 13.5.3 section 9.3.2
- For content filters, inspect false positives and false negatives alongside user feedback, expert review, A/B tests, performance, bias, and legal compliance. section 13.5.3
Evaluating a generated summary
ROUGE uses precision to assess how much of a system-generated summary is relevant, needed, or useful compared with a human reference summary. Because ROUGE relies on surface-level overlap, precision should be combined with other metrics and human judgment to assess meaning and fluency. section 12.2.2
