Evaluating Recurrent Neural Network Explanations