Optimizing annotation resources for natural language de-identification via a game theoretic framework