AI Researchers Adapt Decision Transformers to Help Revive Endangered Heritage Languages

A researcher exploring offline reinforcement learning techniques discovered a potential application for Decision Transformers in heritage language preservation after a colleague working with the Cherokee Nation raised challenges around building tutoring systems for critically endangered languages. The core problem, termed 'extreme data sparsity,' involves languages with fewer than 2,000 fluent speakers, minimal digitized corpora, inconsistent orthography, and only a handful of elder speakers. Standard NLP approaches requiring 10,000-plus parallel sentences are impractical for languages like Ainu or Livonian, which have fewer than 30 fluent speakers worldwide. The proposed solution reframes language learning as a sequential decision-making problem, using Decision Transformers — originally developed by Chen et al. at UC Berkeley — to condition pedagogical decisions on target proficiency outcomes rather than relying on large training datasets. The research highlights both the technical and ethical complexities of deploying AI in culturally sensitive, resource-scarce linguistic communities.