Evaluating Distributional Distortion in Neural Language Modeling