Sheffield's Submission to the AmericasNLP Shared Task on Machine Translation into Indigenous Languages
Edward Gow-Smith, Danae Sánchez Villegas
TL;DR
This work tackles Spanish-to-eleven Indigenous language translation under low-resource conditions by extending and ensembling variants of the NLLB-200 model and training on diverse data sources, including organizers’ data, historical corpora, and backtranslations. The team conducts multilingual fine-tuning, explores various subsystems (single best, per-language Best, and ensembles), and evaluates with chrF. Results show the proposed approaches yield the highest average chrF on the test set, with top performance in four languages and strong improvements overall, driven by data diversity, transfer learning, and backtranslation for data-rich languages. The findings demonstrate that multilingual training and model ensembling can substantially improve translation quality in low-resource, morphologically complex indigenous languages, suggesting practical applicability for indigenous language preservation and communication workflows.
Abstract
In this paper we describe the University of Sheffield's submission to the AmericasNLP 2023 Shared Task on Machine Translation into Indigenous Languages which comprises the translation from Spanish to eleven indigenous languages. Our approach consists of extending, training, and ensembling different variations of NLLB-200. We use data provided by the organizers and data from various other sources such as constitutions, handbooks, news articles, and backtranslations generated from monolingual data. On the dev set, our best submission outperforms the baseline by 11% average chrF across all languages, with substantial improvements particularly for Aymara, Guarani and Quechua. On the test set, we achieve the highest average chrF of all the submissions, we rank first in four of the eleven languages, and at least one of our submissions ranks in the top 3 for all languages.