On Sample Based Explanation Methods for NLP:Efficiency, Faithfulness, and Semantic Evaluation