Why might F1 score be a better choice than raw accuracy for evaluating entity extraction?

**F1 score** is the harmonic mean of precision and recall. It gives you a single number that balances both directions of error: false positives (precision) and false negatives (recall). Precision asks: of all the entities the model predicted, how many were actually correct? A model with high precision rarely labels something as an entity when it is not one. Recall asks: of all the real entities in the data, how many did the model find? A model with high recall catches most of the actual entities, even if it also predicts some false ones. $$F_1 = \frac{2 \cdot \text{precision} \cdot \text{recall}}{\text{precision} + \text{recall}}$$ Raw accuracy counts the fraction of all predictions that are correct. For entity extraction, this is misleading because most tokens are not entities. If only 5% of tokens carry entity labels, a model that always predicts 'no entity' gets 95% accuracy. Its F1 is zero because it has zero recall. F1 is preferred because it forces you to actually find entities (recall) while keeping your predictions clean (precision). Neither direction is hidden.

**F1** is the harmonic mean of precision and recall, and it is preferred over raw accuracy for entity extraction because accuracy hides poor performance on rare classes. $$F_1 = \frac{2 \cdot \text{precision} \cdot \text{recall}}{\text{precision} + \text{recall}}$$ Precision measures the correctness of positive predictions: of everything the model labeled as an entity, what fraction was actually an entity? Recall measures completeness: of all the real entities in the test set, what fraction did the model find? The harmonic mean is strict. Unlike the arithmetic mean, it cannot be high if either component is low. If precision is 1.0 but recall is 0.0, F1 is 0.0, not 0.5. This property is exactly what makes F1 useful for imbalanced tasks. Entity extraction is inherently imbalanced. In a typical named entity recognition dataset, 90% or more of the tokens are not entities. A model that never predicts an entity gets accuracy above 90% but F1 of exactly zero. Accuracy is dominated by the easy majority class (non-entities) and completely hides the model's failure on the minority class (actual entities). The practical workflow: report F1 as the primary metric for entity extraction, and report precision and recall separately to diagnose whether the model is too aggressive (low precision, high recall) or too conservative (high precision, low recall).