NüshuRescue hits 48.69% accuracy on endangered script

The paper, presented at COLING 2025 in Abu Dhabi as part of the Proceedings of the 31st International Conference on Computational Linguistics, tackles a problem the authors describe as typically labor-intensive and costly: languages with very little written material. Nüshu is their example — a script historically used by Yao women in China — and the scale of the challenge is visible in the numbers: a 500-sentence parallel corpus, 35 training examples, and 50 sentences held back for testing.

The team's approach rests on two supporting pieces. NCGold, the 500-sentence Nüshu-Chinese parallel corpus, is described as the first publicly available dataset of its kind, and NCSilver, a set of 98 newly translated modern Chinese sentences of varying lengths, extends the target corpus. Alongside the GPT-4-Turbo work the authors built FastText-based and Seq2Seq models, so the release covers more than a single method. All datasets and code were made public.

The reported 48.69% accuracy is a starting point rather than a finished translation tool, and the evaluation rests on one small held-out set, so how far this carries to other endangered languages is not yet settled. The near-term test is whether other researchers pick up the released corpus and code, since the paper's claim of minimizing human input depends on that reuse.