Pretraining Language Models with Subword Regularization: An Empirical Study of BPE Dropout in Low-Resource NLP

ArXi:2605.13436v1 Announce Type: cross Subword regularization methods such as BPE dropout are typically applied only during fine-tuning, while pre