Aashish Dhawan, Christopher Driggers-Ellis, Christan Grant, Daisy Zhe Wang
LoResMT 2026 · EACL · 2026
Parallel corpora for low-resource Indigenous languages are scarce, which limits neural machine translation quality. We augment the available data with synthetic sentence pairs generated by a high-capacity multilingual translation model, and fine-tune mBART-50 on both curated-only and synthetically augmented data, evaluating with chrF++. We pair this with language-specific preprocessing — orthographic normalization and noise-aware filtering. Results show consistent gains for Guaraní–Spanish and Quechua–Spanish, while experiments on Aymara expose the limits of generic preprocessing for highly agglutinative languages, motivating morphology-aware handling.
We tackle data scarcity from two directions. First, we generate synthetic parallel sentence pairs with a high-capacity multilingual translation model to expand the curated corpora. Second, we apply language-specific preprocessing — orthographic normalization to reduce spurious surface variation, and noise-aware filtering to keep augmentation from injecting low-quality pairs.
We fine-tune mBART-50 on both curated-only and augmented data and compare with chrF++, isolating the contribution of synthetic data and of each preprocessing step.
Augmentation and normalization yield consistent improvements for Guaraní–Spanish and Quechua–Spanish. Aymara is the instructive exception: its highly agglutinative morphology blunts generic preprocessing, pointing toward morphology-aware normalization as the next step.
On the AmericasNLP development sets, adding forward-translated synthetic data improves chrF++ across all three languages — largest for Aymara (+3.97), followed by Guaraní (+2.00) and Quechua (+0.71) — competitive with reported shared-task systems.
Back to Aashish Dhawan's site