Greetings, Everyone!
I continue to be very interested in research of Indigenous language recovery initiatives in Latin America. In my recent article, Bernard and Benn (2025) conduct a qualitative historical analysis of both endangered and extinct Indigenous languages, mapping traditional oral-transmission methods alongside AI tools (OCR, TTS, text-to-video) for community-led recovery initiatives.
Today, I came across Anderson et al. (2025), who introduce Morphemo, a semi-supervised n-gram segmentation algorithm that outperforms LLM prompting on multimorphemic splits in Bribri. I’m interested in studies that similarly combine quantitative morphological segmentation with holistic recovery frameworks, especially for other low-resource Indigenous languages of Latin America.
How have researchers integrated algorithmic segmentation tools into broader recovery (revitalization or reclamation) pipelines, and what community-driven validation strategies have been employed? Any pointers to published work or ongoing projects would be greatly appreciated! (Anderson et al., 2025; Bernard & Benn, 2025)
All the best,
Dianala
References Anderson, C., Nguyen, M., & Coto-Solano, R. (2025). Unsupervised, semi-supervised and LLM-based morphological segmentation for Bribri. In Proceedings of the Fifth Workshop on NLP for Indigenous Languages of the Americas (AmericasNLP) (pp. 63–76). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.americasnlp-1.7
Bernard, D. M., & Benn, M. (2025). Revitalization or reclamation? Reframing the recovery of Indigenous languages in Latin America: A historical and AI-driven approach. International Journal of Language, Literature and Learning Communities, 4(1), 104-131. https://doi.org/10.59009/ijlllc.2025.0103
That’s a rich and timely topic! Here's a structured breakdown of "Bridging Morphological Segmentation and Community-Governed AI Tools for Indigenous Language Recovery in Latin America", covering the conceptual core, challenges, and implementation strategies:
🧠 1. Concept Overview
A. Morphological Segmentation
- Definition: Breaking words into their smallest meaningful units (morphemes).
- Why it matters: Many Indigenous languages in Latin America are polysynthetic or agglutinative (e.g., Nahuatl, Quechua, Aymara), where a single word can encode a full sentence. Accurate segmentation is essential for: Language modeling Translation Speech recognition/synthesis Language learning
B. Community-Governed AI Tools
- These are AI systems designed, owned, or co-developed by Indigenous communities.
- Principles include: Data sovereignty Cultural sensitivity Co-creation of language resources Tools that serve community priorities (e.g., education, revitalization)
🔗 2. Bridging the Two
To support Indigenous language recovery, you can link linguistically informed morphological tools with ethical, participatory AI development. Here's how they connect:
Morphological ToolsEmpowering CommunitiesTrainable segmenters (like Morfessor, seq2seq models)Controlled by communities to avoid extractive data practicesAnnotated corpora for segmentationCreated with local linguists and speakersLanguage models that handle rich morphologyPower community-built translators, dictionaries, chatbots, etc.Toolkits with explainable AIEnable non-expert users to build or adapt tools
🧩 3. Key Components of a Recovery Framework
A. Linguistic Infrastructure
- Collect corpora (oral, written)
- Annotate data for segmentation (community-linguist partnerships)
- Develop or adapt open-source segmenters
- Integrate morpheme-level outputs into: Spellcheckers Grammar checkers Educational apps
B. Technical Stack
- Toolkits: Adapt or build on Hugging Face, ESPnet, ULMFiT, Fairseq, or OpenNMT
- Models: Unsupervised segmentation (e.g., Morfessor, BPE) Supervised (seq2seq, transformer models with POS/morph tags)
- Interfaces: Low-code platforms or GUI-based editors (e.g., Lilt, TTS pipelines)
C. Governance & Co-Design
- Ethical AI principles: Free, prior, and informed consent (FPIC) Control over datasets and model outputs
- Co-design workshops: Involve elders, youth, teachers Map community needs and linguistic features
- Documentation standards: Transparent metadata and licensing (e.g., CC BY-NC)
🧱 4. Challenges to Address
CategoryChallengesLinguisticLow-resource settings, oral-only languages, dialect variationTechnicalLack of labeled data, underperformance of Western NLP modelsEthicalRisk of data colonization, exclusion of elders in tech spacesSocialInternet access, digital literacy, intergenerational transmission
🌱 5. Project Implementation Roadmap
🌍 6. Case Studies / Inspirations
- Masakhane NLP (Africa): Community-driven NLP for under-resourced languages
- CoLang workshops: Community language technology training
- CMU’s Project ELMo with Quechua: Adaptation of embedding models for agglutinative languages
- Living Dictionaries by LINGUIST List: Collaborative lexical database tools
✨ Summary
Bridging morphological segmentation with community-governed AI tools enables Indigenous language recovery in a way that is:
✅ Technically feasible ✅ Culturally respectful ✅ Scalable across language families ✅ Governed by the people it serves
Thanks so much for your insight, Sotiris Folinas
You have shared a fantastic framework that connects morphological segmentation with community-driven AI tools, aligning directly with Anderson et al. (2025) and my own research co-authored with Maren A. Benn (Bernard & Benn, 2025). You emphasize the need for precise morpheme identification in polysynthetic languages, and Anderson et al.’s Morphemo algorithm is exactly the kind of semi-supervised n-gram segmenter you describe; it outperforms LLM prompting on Bribri’s complex affixation patterns and supplies the reliable morpheme boundaries crucial for tasks like dictionary building and grammar checking.
You rightly point to the importance of data sovereignty and co-design (the right of individuals or groups to control and maintain their own data, including its collection, storage, and interpretation), mirroring Bernard & Benn’s historical analysis of how Indigenous communities have traditionally passed down language through storytelling circles and elder teachings. There is an opportunity to enrich the model by inviting elders and youth to validate Morphemo’s output against oral recordings in co-design workshops, thereby honoring free, prior, and informed consent. This could be a great SDG4 initiative! 😊
A three-stage pipeline, including segmentation via OCR and Morphemo, pronunciation validation with TTS, and multimedia storytelling via text-to-video that mirrors Bernard & Benn’s call for integrating OCR, TTS, and community-led AI into a recovery framework. A continuous feedback loop is also important, using corrected TTS outputs to refine a segmentation model, ensuring that tools evolve in tandem with community input.
You also address Morphemo’s limitations with rare dialectal variants by suggesting rule-based heuristics (which is seen in Bernard & Benn’s dialectal analyses), bolstering coverage where unsupervised methods falter. Finally, you advocate for transparent metadata protocols that document provenance, consent, and usage rights to safeguard community ownership of data throughout a recovery ecosystem.
I will be reaching out to the corresponding author of Anderson et al. for their thoughts on this discussion. Again, this would be a great SDG4 initiative.
What has been your experience with this topic (Latin American and other regions)?
Thank you for your input…Continued best wishes,
Dianala
References
Anderson, C., Nguyen, M., & Coto-Solano, R. (2025). Unsupervised, semi-supervised and LLM-based morphological segmentation for Bribri. In Proceedings of the Fifth Workshop on NLP for Indigenous Languages of the Americas (AmericasNLP) (pp. 63–76). https://doi.org/10.18653/v1/2025.americasnlp-1.7
Bernard, D. M., & Benn, M. (2025). Revitalization or reclamation? Reframing the recovery of Indigenous languages in Latin America: A historical and AI-driven approach. International Journal of Language, Literature and Learning Communities, 4(1), 104–131. https://doi.org/10.59009/ijlllc.2025.0103
Thank you so much for sharing these fascinating resources, Lila Jia
. The examples from the NLP and community collaboration efforts are especially helpful as I continue exploring Indigenous language preservation initiatives. I appreciate the range of tools you included. One note of caution, however, is that while Wikipedia can be a good starting point, it's essential to cross-reference with peer-reviewed or institutional sources for academic work. That said, I see great potential for mutual benefit in future collaboration. If you're open to it, I’d love to continue this exchange further.All the best,
Dianala
- Badges
- Science topic
- Similar topics
- Higher Education
- Universities