Kumatigi: Quality-Driven Data Augmentation for Low-Resource Machine Translation
Cheick Tidiani Cissé
Findings of the Association for Computational Linguistics: ACL 2026, pp. 24433–24446
Built and released a 67k quality-scored French↔Bambara corpus, showing that filtering for quality beats scaling raw volume in extremely low-resource translation. +3–4 BLEU over baselines, confirmed by human evaluation.