COLING 2025: AI Framework Unveiled for Restoring the Endangered Nüshu Language Using LLMs

At COLING 2025, the NüshuRescue framework was unveiled, which trains large language models (LLMs) on small amounts of data to restore endangered languages.

Key Technologies and Results

• Data Efficiency: Using GPT-4-Turbo, the framework achieved 48.69% translation accuracy on 50 test sentences with only 35 short examples, without prior exposure to Nüshu.
• Dataset Construction: Developed NCGold (500 sentences), a Nüshu-Chinese parallel corpus, and generated NCSilver, a set of 98 new modern Chinese translation sentences based on it.
• Auxiliary Models: FastText-based and Seq2Seq models were also developed to support the research.

Significance

This framework serves as a tool to accelerate the reconstruction of endangered languages while minimizing human input. All datasets and code have been made publicly available via GitHub.