CL2GEC: A Multi-Discipline Benchmark for Continual Learning in Chinese Literature Grammatical Error Correction

ICLR 2026 Conference Submission25228 Authors

Keywords: Chinese Grammatical Error Correction;Benchmark Evaluation;Continual Learning;Large Language Models

TL;DR: CL2GEC is a new benchmark for Chinese academic GEC across 10 disciplines; results show that regularization-based continual learning significantly outperforms replay and sequential tuning in both grammatical accuracy and knowledge retention.

Abstract: The growing demand for automated writing assistance in scientific domains highlights the need for robust Chinese Grammatical Error Correction (CGEC) systems that can adapt across disciplines. However, existing CGEC research lacks dedicated benchmarks for academic writing and overlooks continual learning as a solution to handle domain-specific variation. To fill this gap, we introduce CL2 GEC, a Continual Learning benchmark for Chinese Literature Grammatical Error Correction, designed to evaluate adaptive CGEC across multiple academic fields. Our benchmark includes 10,000 human-annotated sentences spanning 10 disciplines, each exhibiting distinct linguistic styles and error patterns. We evaluate large language models under sequential tuning, parameter-efficient adaptation, and representative continual learning strategies, using both standard GEC metrics and continual learning metrics adapted to task-level variation. Experimental results show that regularization-based continual learning methods, such as OGD and GEM, outperform replay-based and sequential approaches in both grammatical accuracy and knowledge retention. These findings underscore the feasibility and importance of integrating continual learning into CGEC and position our benchmark as a foundation for future research on adaptive scientific writing assistance.

Primary Area: datasets and benchmarks

Submission Number: 25228