F1 score compresses precision and recall into one classification or retrieval metric. Precision measures how many predicted positives are correct; recall measures how many actual positives are found. Because it uses their harmonic mean, improving both raises the score, while weakness in either remains visible. section 4.2.1 section 12.7.2
Toxic-comment moderation
A toxicity detector needs both precision and recall. Low precision sends too many harmless comments to moderators or hides legitimate posts; low recall lets more toxic comments through. F1 provides one benchmark for comparing models, but the underlying precision and recall still explain the operational trade-off. section 4.2.1
Use F1 to evaluate prediction systems at the level that matters:
- Compare classifiers with one combined score when both missed positives and incorrect positive predictions matter. section 4.2.1
- Evaluate intent-based chatbots with binary or generalized multiclass F1. section 12.7.2
- For multi-label chatbots, average the individual intent F1 scores and inspect each intent separately to find weak points. section 12.7.2
F1 is not a complete evaluation picture. Inspect precision and recall alongside it, especially when deploying systems that alert moderators, hide content, or classify user intent. Retrieval-based chatbot evaluation uses different metrics, such as exact match, rather than automatically applying F1. section 4.2.1 section 12.7.2
