Pruning Literals for Highly Efficient Explainability at Word Level