Tag
This paper proposes a modular approach for adapting pretrained language models to low-resource languages by freezing embeddings and tuning the rest, showing improvements on NLU tasks for Scottish Gaelic, Irish, and Quechua.