Regularization and nonlinearities for neural language models: when are they needed?