Vocabulary expansion tailored to non-Latin scripts dramatically improves language model performance on low-resource African languages, with gains transferring to related languages without additional training.
This paper improves multilingual AI models for African languages written in Ge'ez script (like Amharic and Tigrinya) by expanding their vocabulary with script-specific words. The researchers created VEXMLM by adding 30,000 Ge'ez-script tokens to XLM-R and training it on curated text, achieving major performance gains—for example, improving question-answering accuracy from 66% to 87% on Amharic.