Edit-level Majority Voting Mitigates Over-Correction in LLM-based Grammatical Error Correction
Researchers have developed a new method to address the over-correction problem in large language models used for grammatical error correction. Their training-free inference technique involves generating multiple correction candidates from a single model and then applying an edit-level majority vote. This approach has shown superior performance compared to standard decoding methods across nine diverse language benchmarks, while also maintaining consistent quality regardless of the input prompts. AI
IMPACT This novel method offers a practical way to enhance the accuracy of LLM-based grammar correction tools without requiring additional training.