True Negatives for Named Entity Recognition

By definition, true negative is non-entity tokens in the text, agreed by gold standard. So;

TN: 5 = {announced, the, release, of}

But, in practice the number of true negative tokens are quite high and this heavily dominates the metrics, and any measure using TN (like accuracy) becomes highly controversial. Instead, preferred better metrics:

• Recall: TP/(TP+FN) - coverage in all correct entities, and
• Precision: TP/(TP+FP) - how reliable for a positive result
• F1 Score (harmonic mean of precision and recall) can be used as a balanced metrics.