A computer system can recognise an expression such as Territorio_Madre_Tierra, distinguish it from Autonomía_Territorial, and connect it with Memoria_Histórica. It can classify an intent, extract entities, and produce an explanation in Spanish.
The most difficult part begins afterwards.
What happens when a language bound to territory, spirituality, and community life becomes a labelled corpus? Who decides whether an interpretation truly preserves its meaning? What rights accompany phrases when they move from an Indigenous reserve to a digital repository and, from there, may be incorporated into other artificial intelligence systems?
An article published on 21 January 2026 in the Colombian journal Entramado presents a proof of concept for interpreting expressions in Nasa Yuwe using natural language processing. The study was developed by Darcyn Frynet Embus-Quina, Andrés Alexis Elago-Quitumbo, and Angela María Rodríguez-Vivas, from Corporación Universitaria Comfacauca, and was identified through Dialnet’s academic radar.
Translation Is Not Word Substitution
The project focuses on the variety of Nasa Yuwe spoken in the Indigenous reserve of San Lorenzo de Caldono, in Colombia’s Cauca department. This delimitation reflects the dialectal, phonetic, and orthographic differences between Nasa territories. Rather than bringing together expressions from different places under an apparently uniform form, the team chose to work with one specific territorial variety.
Nor is the aim to build a conventional translator that replaces each Nasa Yuwe word with a Spanish equivalent. The system attempts to identify communicative intent and recognise the cultural, historical, and territorial entities present in each expression.
The categories used to annotate the corpus include:
Territorio_Madre_TierraAutonomía_TerritorialMemoria_HistóricaParticipación_ComunitariaDefensa_TerritorioResistencia_PacíficaCosmovisión
These labels allow the programme to produce more than a lexical equivalent. Its output may include an explanation of the meaning attributed to the expression and its relationship with territory, history, or collective organisation.
This technical decision contains an important anthropological insight. Meaning does not reside solely in a sequence of words. It also depends on relationships with the land, experiences of violence, forms of authority, community practices, and ways of understanding life.
The article attempts to prevent those relationships from disappearing in translation. That is its principal contribution.
A Language with Extremely Limited Digital Resources
Current machine translation systems benefit from enormous quantities of parallel texts, dictionaries, recordings, and annotated documents. These resources are distributed in profoundly unequal ways. English, Spanish, and French have accumulated linguistic infrastructures over decades, while many Indigenous languages are barely present in the digital environments used to train models.
The study describes Nasa Yuwe as a language with extremely limited digital resources. To build the prototype, the team began with 50 original sentences, contributed through participatory consultations by members of the Tejido del Nasa Yuwe.
Naming this collective avoids presenting the community as an anonymous source of data. According to the article, the Tejido has worked since 2005 on designing and implementing curricula to strengthen written and spoken use of the language across different educational sites in San Lorenzo de Caldono. It also organises workshops, produces teaching materials, and brings together teachers who speak the language and those who do not.
The selected sentences come from community spaces, cultural events, and meetings with government institutions or representatives of the state. The article itself defines the corpus as a set of expressions related to political debates. It is therefore not a neutral inventory of greetings and everyday objects. It includes ways of speaking about territory, autonomy, justice, memory, and relations with the state.
Starting from the 50 sentences, the team manually augmented the data by dividing and reorganising them. The training corpus consequently reached 167 examples, containing 1,933 tokens, 19 intents, and 29 cultural entities.
The model was built with Rasa and its DIET component, a transformer-based architecture capable of classifying intents and extracting entities. The annotations were carried out manually. The process preserved Nasa Yuwe diacritics and phonetic symbols, although adjustments were also needed to prevent these characters from interfering with the syntax used by Rasa.
Here, engineering involves more than choosing an algorithm. It must decide what to preserve, what to normalise, and what to turn into a category. Each of these operations can affect the way in which the language becomes legible to the machine.
A Perfect Score Does Not Learn a Worldview
The published results are promising for a proof of concept on this scale.
Across 144 evaluated samples, intent classification achieved an F1 score of 100 per cent. In the extraction of 955 annotated entities, the system reached macro precision of 91.1 per cent, recall of 95.3 per cent, and an overall F1 score of 92.8 per cent.
A high accuracy score does not, however, mean that the model has learnt a worldview.
It means that it has successfully recognised the categories defined within a small, carefully prepared dataset related to the training material. External evaluations and a much larger dataset would be needed to establish how the system responds to spontaneous speech, new speakers, or expressions not derived from the initial corpus.
The article itself points to confusion between Gobierno_Propio and Justicia_Propia as an example of closely related concepts that cannot be separated by stable semantic boundaries within the Nasa worldview. Territorio_Ancestral_Nasa and Derecho_Territorial likewise appear among the entities with the weakest performance.
These scores remain high, but the differences between precision and recall show that some categories central to political and territorial life are also difficult to turn into computational compartments.
The problem is not merely a lack of data. A taxonomy may force apart dimensions that appear interconnected in community discourse. The machine needs boundaries in order to classify; a worldview may organise meaning through relationships, overlaps, and contexts that cannot be stabilised so easily.
A Language Contains Territorial Differences
Working with the San Lorenzo de Caldono variety is a reasonable methodological decision. It avoids indiscriminately mixing expressions and recognises that Nasa Yuwe is not a uniform form.
The same delimitation raises a question for the future. If the prototype is expanded and presented as a general interpreter of the language, what will happen to the varieties spoken in other territories? An educational tool or automated system may unintentionally turn the form with the greatest amount of data into the form considered correct.
This risk is not unique to this project. Digital language systems tend to reinforce what they can measure most easily. The best-documented variety achieves stronger results, is incorporated into more applications, and may eventually acquire a technical authority that is mistaken for linguistic authority.
A comparison with the work of José Robinson Yaqui Pete makes this internal diversity visible. Yaqui Pete is a systems engineer, teacher, and member of the Nasa people. He is developing a language-strengthening project in the ancestral territory of Llano Buco, Huila, a Nasa territory distinct from Caldono.
His initiative uses digital media to recover vocabulary, create materials, strengthen bilingual education, and bring the language closer to children and young people. Its horizon includes applications, digital dictionaries, translators, videos, podcasts, and virtual communities, but it begins with needs formulated within the territory itself.
The comparison does not require choosing between university research and community-led language activism. It shows that there are different Nasa experiences, with their own territories, educational processes, and priorities. A responsible infrastructure should be able to recognise this plurality without turning one variety into the measure of all the others.
Who Validates the Interpretation?
The decisive point in the article appears in its discussion and proposed future work.
Final cultural validation had not been completed at the time of publication. The team states that ultimate authority over the validity of the interpreter belongs to the Indigenous community and reports that it is working with the educational community of the Madre Laura school in Caldono. It also proposes expanding the corpus and developing an educational tool in Nasa Yuwe for natural sciences.
This sequence matters.
In a conventional system, quality can be evaluated by comparing an automated output with a translation regarded as correct. In a tool that seeks to interpret territory, memory, justice, or spirituality, authority cannot remain solely in the hands of a confusion matrix.
Cultural accuracy requires another form of validation. It must allow speakers and community institutions to decide:
- whether the categories are appropriate;
- whether an explanation in Spanish preserves or distorts the meaning;
- which territorial differences should be represented;
- which expressions may circulate outside the community;
- which forms of knowledge should not be incorporated into the system;
- who may correct, expand, or remove information;
- which educational, institutional, or commercial uses are acceptable.
Participation should not be limited to providing sentences and reviewing the finished product. It also concerns the design of categories, corpus documentation, access permissions, and the model’s future uses.
What the Repository Does Not Make Visible
The academic article explicitly carries a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence. In the associated repository, however, there is no separate licence for the code and the data, nor a public note detailing attribution to the Tejido del Nasa Yuwe, permitted uses, or procedures for correcting and removing expressions.
This does not allow us to conclude that the research was conducted without agreements. The article describes participatory consultations, acknowledges the Tejido del Nasa Yuwe, and proposes continued collaboration with Caldono’s educational community. What remains unclear in the public documentation is how those relationships are carried into the later stages of the data lifecycle.
The central issue is not whether the corpus is open or closed. It is who can decide on its future uses, who must be acknowledged, which applications may be derived from it, and what capacity the community retains to modify or withdraw materials.
A repository can make data visible without making its governance visible.
From a Visible Corpus to a Governed Corpus
A governed corpus does not have to remain inaccessible. It can be shared for educational or research purposes while still maintaining collective conditions over its circulation.
The CARE Principles for Indigenous Data Governance offer a useful framework. Their four pillars are:
- Collective Benefit
- Authority to Control
- Responsibility
- Ethics
CARE does not replace the FAIR Principles — findable, accessible, interoperable, and reusable — but introduces questions that technical openness cannot resolve. For whom does the data create value? Who retains authority over it? What responsibilities do those who store and reuse it have? Which uses are ethically acceptable from the perspective of the community involved?
Applied to the Nasa Yuwe corpus, this approach would mean that the community did not appear only as the source of the data. It would have a recognised capacity to decide on access, attribution, expansion, corrections, and derived applications.
Local Contexts offers another complementary tool through its Traditional Knowledge Labels, known as TK Labels. These labels allow metadata to carry information about provenance, cultural authority, and community protocols. They include labels for attribution, culturally sensitive or sacred materials, gender-based or seasonal restrictions, non-commercial use, community use only, and openness to collaboration.
TK Labels are not legal licences, nor do they technically prevent a model from being trained. They operate as informational and governance instruments. They make visible rules and responsibilities that conventional repositories often omit.
A constructive proposal for the project could combine:
- separate and appropriate licences for the code and the data;
- explicit attribution to the Tejido del Nasa Yuwe and the participating community institutions;
- protocols governing educational, commercial, and institutional uses;
- Local Contexts metadata defined by the community itself;
- mechanisms for correcting or removing expressions;
- documentation of territorial varieties;
- a community body with effective authority over future expansions.
Language revitalisation does not require turning an entire language into a universally reusable resource.
Language Does Not End in the Model
Artificial intelligence can contribute to educational materials, assistants, writing tools, and consultation systems for languages that major platforms barely consider. It may help younger generations find new spaces in which to read, write, and listen to their language.
It may also establish one variety as the standard, freeze contested categories, separate expressions from their territory, or turn collective knowledge into a resource available to outside institutions.
These possibilities are not mutually exclusive. They coexist within the same project.
The Nasa Yuwe interpreter makes visible the difference between including a language in AI and building AI under the conditions set by those who speak it.
The first task requires data, annotations, and models.
The second requires community authority, the right to correct, the ability to withdraw information, recognition of territorial diversity, and control over future uses.
A language can enter a machine without ceasing to belong to a history and a territory. Technology will remain a tool for revitalisation for as long as it stays subordinate to those who sustain the language and does not seek to become its new owner.
References
Embus-Quina, D. F., Elago-Quitumbo, A. A. and Rodríguez-Vivas, A. M. (2026). Intérprete contextual de expresiones en lengua Nasa Yuwe a español usando técnicas de Procesamiento de Lenguaje Natural. Entramado, 22(1), 1–21. https://doi.org/10.18041/1900-3803/entramado.1.13152
Dialnet. (2026). Bibliographic record and full text of the article.
Embus-Quina, D. F. (n.d.). Nasa Yuwe prototype repository. GitHub.
Global Indigenous Data Alliance. (n.d.). CARE Principles for Indigenous Data Governance.
Carroll, S. R., Garba, I., Figueroa-Rodríguez, O. L. et al. (2020). The CARE Principles for Indigenous Data Governance. Data Science Journal, 19, 43. https://doi.org/10.5334/dsj-2020-043
Local Contexts. (n.d.). Traditional Knowledge Labels.
Álvarez Malvido, M. and Yaqui Pete, J. R. (2025, 14 May). Tecnología y memoria colectiva para fortalecer el nasa yuwe: el proyecto de José Yaqui en Colombia. Global Voices / Rising Voices.
Note on Sources
Dialnet is used as an Ibero-American academic radar. The date, metrics, licence, corpus composition, and problematic categories have been checked against the full text of the article. Information about the repository is limited to what was publicly visible at the time of consultation and does not rule out the existence of unpublished community agreements.