fairseq
fairseq copied to clipboard
Did XLM-R applied subword regularization?
Looking at "MultilingualMaskedLMTask" code, dictionaries seems to be required to setup this task. To build the dictionaries, we require to preprocess the sentence pieces upfront. Preprocessing raw text upfront doesn't allows regularization noise. So the code doesn't seem to apply regularization. Is it that XLM-R doesn't apply subword regularization?