Explaining Predictions of Non-Linear Classifiers in NLP