Multilingual transformer models like mBERT and XLM-RoBERTa have obtained improvements for many NLP tasks on a variety of languages. However, works also showed that results from high-resource languages could not be transferred to realistic, low-resource scenarios. In this work, we study in performance for different amounts of available resources for the African languages Hausa, isiXhosa and Yor\\`ub\\'a on both NER and topic. We show that in combination with transfer learning or distant, these models can achieve with as little as 10 or 100 labeled the same performance as baselines with much more supervised training. However, we also find settings where this does not hold. Our discussions additional experiments on assumptions such as time and hardware highlight challenges and opportunities in low-resource learning.
No takes yet. Share an insight, caveat, or question.
Hedderich et al. (2020) studied this question.