Expanding the Lexicon of Ge'ez Based African Languages: A Comparative Study of Amharic and Tigrinya
A Ge’ez-script vocabulary extension gave XLM-R a large lift on Amharic and Tigrinya tasks.
The paper introduces VEXMLM, an XLM-R variant expanded with 30,000 Ge’ez-script subwords trained from curated Amharic and Tigrinya corpora. It reports 87.0 EM / 90.0 F1 on Amharic/Tigrinya QA, ahead of XLM-R and Glot500 in the abstract’s comparison. Sentiment accuracy rises to 80.0%, while OOV-token entity accuracy on NER increases from 81.4% to 94.3% across the subset where that analysis was possible. The authors also evaluate transfer to 17 additional low-resource African languages. ArXiv · AI/CL/LG's note
The paper introduces VEXMLM, an XLM-R variant expanded with 30,000 Ge’ez-script subwords trained from curated Amharic and Tigrinya corpora. It reports 87.0 EM / 90.0 F1 on Amharic/Tigrinya QA, ahead of XLM-R and Glot500 in the abstract’s comparison. Sentiment accuracy rises to 80.0%, while OOV-token entity accuracy on NER increases from 81.4% to 94.3% across the subset where that analysis was possible. The authors also evaluate transfer to 17 additional low-resource African languages. ArXiv · AI/CL/LG's note
score 4