Expanding the Lexicon of Ge'ez Based African Languages: A Comparative Study of Amharic and Tigrinya
VEXMLM improves multilingual models for low-resource Ge'ez languages like Amharic and Tigrinya by combining language-specific tokenization with 30,000 new Ge'ez subword tokens. This approach, paired with mean initialization and a two-stage training process, significantly boosts performance on tasks like entity recognition and reduces out-of-vocabulary errors.