Explaining text classifiers through progressive neighborhood approximation with realistic samples