Creating Grammar Teaching Material for Endangered Languages with Hybrid Grammar Induction

Explicit grammar teaching plays a central role in endangered language revitalization, yet creating grammar lessons is labor-intensive and is typically assigned to already overwhelmed teachers. We introduce HYGRAM, a Hybrid Grammar Induction method for grammar induction from sparse data that integrates expert linguistic abstractions, typological priors, Bayesian inference, and constrained LLM reasoning. This method is deployed in a teacherfacing software platform designed for languages with minimal digital footprint. Given as little as a corpus elicited within approximately 10 hours of fieldwork and any available descriptive documents, HYGRAM produces topic-specific, structured grammar lessons designed for classroom use. We evaluate the system on six typologically diverse endangered languages with expert linguists, who rate outputs for similarity to expert-authored materials, linguistic quality, and pedagogical usefulness. Results show consistently high similarity, reliable quality once a modest data threshold is reached, and strong recommendations for teacher use (mean 7.7/9). Community feedback from Vanuatu-based language practitioners further indicates high perceived relevance for local revitalization efforts. Our findings demonstrate that controlled, hybrid grammar induction can support practical grammar teaching in extremely low-resource settings.