Edit-level Majority Voting Mitigates Over-Correction in LLM-based Grammatical Error Correction

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Advanced, short

Summary

A new training-free inference method, "edit-level majority voting," addresses the over-correction issue prevalent in large language model (LLM)-based grammatical error correction (GEC). Proposed by Takumi Goto, Yusuke Sakai, and Taro Watanabe, this technique operates by performing majority voting over multiple correction candidates generated by a single LLM. Crucially, it requires no model modifications or additional training. Evaluated across nine benchmarks encompassing English, Czech, German, Ukrainian, Korean, Hindi, and Romanian, the method demonstrates superior performance compared to both greedy and MBR decoding in most scenarios. Furthermore, it ensures stable correction quality, regardless of the instruction prompts utilized. The researchers have also released two repositories to support GEC datasets and LLM inference.

Key takeaway

For NLP Engineers developing grammatical error correction systems with large language models, this research offers a critical solution to the common over-correction problem. You should consider implementing the proposed training-free edit-level majority voting method. This approach enhances correction quality and stability across diverse languages, outperforming traditional decoding methods, all without requiring expensive model modifications or additional training. It provides a direct path to more reliable GEC outputs.

Key insights

LLM over-correction in GEC is mitigated by training-free edit-level majority voting across multiple candidates.

Principles

• Majority voting enhances GEC reliability.
• Stable correction is possible without prompt tuning.
• Training-free methods can improve LLM inference.

Method

The method involves generating multiple grammatical correction candidates from a single LLM, then applying edit-level majority voting across these candidates to produce a final, less over-corrected output.

In practice

• Implement edit-level voting for GEC tasks.
• Utilize existing LLMs without fine-tuning.
• Explore provided GEC datasets and inference tools.

Topics

• Grammatical Error Correction
• Large Language Models
• Over-correction Mitigation
• Edit-level Majority Voting
• Inference Methods
• Multilingual NLP

Best for: Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist, NLP Engineer

Related on AIssential

• How I Improved Speech-to-Text Accuracy — A two-pass LLM post-processing method effectively corrects common speech-to-text transcription errors.
• Reliable Chain-of-Thought via Prefix Consistency — Truncating and regenerating LLM Chain-of-Thought traces reveals a reliability signal for improved answer aggregation.
• Encode Errors: Representational Retrieval of In-Context Demonstrations for Multilingual Grammatical Error Correction — LLMs' internal states contain extractable grammatical error representations (GER) that improve in-context learning for GEC.
• Improving Cross-Format Robustness in Language Models with Multi-Format Training — Multi-format training, especially with partial data augmentation, significantly improves LLM robustness to answer format changes.
• Self-Generated Error Training for Token Editing in Diffusion Language Models — Training language model editors on self-generated errors improves performance by aligning training with inference conditions.
• CodeChat-Eval: Evaluating Large Language Models in Multi-Turn Code Refinement Dialogues — LLMs significantly degrade in functional correctness during multi-turn code refinement, especially with semantic or additive changes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.