AmericasNLP Shared Tasks
AmericasNLP is a workshop and shared task series, co-located with major NLP conferences, devoted to Indigenous languages of the Americas. A shared task is a research competition: the organizers release a common dataset and a problem, and teams from around the world build and compare systems against it. Since 2021 the tasks have centered on machine translation between Spanish and Indigenous languages, and more recent editions have added tasks like generating educational material and developing better evaluation metrics. The 2025 edition spanned 14 languages with 11 teams. The 2026 task turned to culturally grounded image captioning.
Context
Good work with Indigenous communities is built on relationships, and relationships take time. The point of the work should be to build things the community actually wants and can use, on terms the community sets. That priority can sit uneasily with another thing research depends on: curiosity. A lot of what makes research move forward is people exploring questions because they are interesting, before anyone can say what they are "for."
Academics may recognize the same tension in the funding landscape, which increasingly rewards translational work with a clear, near-term "economic impact" and leaves less room for open-ended exploration. Insisting that every effort on an Indigenous language prove its usefulness up front would prevent the kind of curiosity-driven work that, over time, produces useful things. It also makes the field less welcoming to newcomers.
The AmericasNLP shared tasks have been a way to balance these priorities. They are a space where people, including newcomers, can explore inside a space the organizers have made safe. That safety exists because the organizers have done the slow relational work, with the communities whose languages and data are in play, that individual participants could not do on their own.
Why it works
- Organizers built relationships. The organizers stand between participants and communities, vetting what data is shared and on what terms, so newcomers can take part without extracting from a community directly.
- Low barrier, high standards. A well-scoped, ethically prepared entry point lets people try working on these languages without already having community ties.
What to watch
- Boundaries need maintaining. The model works only because the organizers actively control what is released.
- Benchmarks flatten. A leaderboard score can obscure what matters to speakers, and the system that "wins" is not necessarily one a community would want to use.