On the Explainability of Natural Language Processing Deep Models