Modular Monolingual Adaptation using Pretrained Language Models

In three linesModular approach to adapt pretrained language models to low-resource languages. Instead of full model retraining, authors replace tokens, freeze corresponding embeddings, and tune the rest. Experiments on Scottish Gaelic, Irish, and Quechua (8.5k instances) show improvements on NLU tasks (mask filling, NER, POS).

Summary generated by Claude — human-verified