Improving Chinese Grammatical Error Correction via Iterative Feedback-Guided Data Evolution
Yuan LI , Zhengtao YU , Ying LI , Hua LAI , Ling DONG , Shengxiang GAO , Cunli MAO , Yuxin HUANG
Chinese Grammatical Error Correction (CGEC) involves detecting and correcting diverse grammatical errors in input text. While parameter optimization improves CGEC performance, current fine-tuning methods assign equal weight to all training samples, leading to inefficient corpus utilization and failing to accommodate the model’s evolving data requirements. In this work, we propose an iterative feedback-guided data evolution framework to improve CGEC performance by leveraging dynamic sample probing and progressive data synthesis. For dynamic sample probing, we design a difficulty-aware selection mechanism to quantify sample value via gradient contributions, thereby dynamically re-weighting instances to prioritize those highly aligned with the model’s evolving state. For progressive data synthesis, we explore a curriculum-driven strategy that adaptively generates complexity-incremental error patterns to achieve dynamic alignment between data distribution and the model’s evolving capacity, thereby facilitating a deeper grasp of semantic structures. Experimental results on multiple benchmarks demonstrate the effectiveness and robustness of our method. Further analysis reveals that the data evolution framework effectively mitigates the misalignment between static data supply and dynamic model requirements, confirming that the sample-model co-evolutionary strategy is pivotal for attaining robust generalization within CGEC tasks.
Chinese grammatical error correction / data evolution / dynamic sample probing / progressive data synthesis
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(22673KB)
