The Design and Evaluation of Neural Attention Mechanisms for Explaining Text Classifiers