Encode Errors: Representational Retrieval of In-Context Demonstrations for Multilingual Grammatical Error Correction
Signal
72
Hype
25
In three linesRetrieval method for in-context demonstrations using Grammatical Error Representations (GER) for multilingual grammatical error correction. On 8B open-source models, results match GPT-4o-mini and Deepseek2.5. For low-resource languages, F₀.₅ scores improve up to 1.20× over baseline.Read source
Your take?
Summary generated by Claude — human-verified