Answers are grounded in this paper and cite their sources.
This work addresses Neural Machine Translation (NMT) for Spanish–Mapudungun, an endangered and digitally under-resourced indigenous language of South America with polysynthetic and agglutinative morphology. We explore transfer learning as a strategy to mitigate data scarcity, initializing Spanish–Mapudungun models from high-resource pairs (Spanish–English and Spanish–Finnish). Using a standardized 10,000-pair corpus, we show that transfer learning markedly improves translation quality in both directions over models trained from scratch. We further analyze whether the linguistic characteristics of the parent language—particularly morphological similarity—affect transfer effectiveness. Results indicate that both parent models achieve similar gains, suggesting that vocabulary coverage and tokenization dominate over morphological resemblance. These findings demonstrate the utility of transfer learning for endangered languages and underscore the need for linguistically informed methods and community involvement in revitalization efforts.
No takes yet. Share an insight, caveat, or question.
Lira et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: