Vocabulary trimming, a common practice in neural machine translation, fails to consistently improve model performance and can even lead to substantial degradation across a wide range of hyperparameter settings.
