Self-Normalized Importance Sampling for Neural Language Modeling