Improving Indigenous Language Machine Translation with Synthetic Data and Language-Specific Preprocessing

TL;DR

This paper tackles MT for low-resource Indigenous languages by combining forward-translated synthetic data with language-specific preprocessing to augment small curated corpora. The authors fine-tune a multilingual mbart-large-50 model using Spanish-pivot synthetic data from MultiScript30k and assess performance primarily with chrF++, reporting consistent gains for Guarani–Spanish and Quechua–Spanish, while Aymara–Spanish shows more modest improvements due to its strong agglutinative morphology. They implement targeted preprocessing per language (Guarani, Quechua, Aymara) and provide detailed experimental results including a Quechua orthographic normalization that substantially boosts chrF++ to 36.6 on development data. The work aligns with AmericasNLP findings that data-centric augmentation with robust multilingual models yields practical gains for Indigenous MT and highlights remaining challenges, notably the need for morphology-aware preprocessing and human evaluation to capture cultural nuance. Overall, the study demonstrates the viability of synthetic data augmentation as a scalable strategy to improve translation quality in data-scarce Indigenous language settings and informs future directions in morphology-aware preprocessing and multimodal augmentation.

Abstract

Low-resource indigenous languages often lack the parallel corpora required for effective neural machine translation (NMT). Synthetic data generation offers a practical strategy for mitigating this limitation in data-scarce settings. In this work, we augment curated parallel datasets for indigenous languages of the Americas with synthetic sentence pairs generated using a high-capacity multilingual translation model. We fine-tune a multilingual mBART model on curated-only and synthetically augmented data and evaluate translation quality using chrF++, the primary metric used in recent AmericasNLP shared tasks for agglutinative languages. We further apply language-specific preprocessing, including orthographic normalization and noise-aware filtering, to reduce corpus artifacts. Experiments on Guarani--Spanish and Quechua--Spanish translation show consistent chrF++ improvements from synthetic data augmentation, while diagnostic experiments on Aymara highlight the limitations of generic preprocessing for highly agglutinative languages.