Elamur Research⌕
☰
home / papers / 2026 / CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards

CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards

Paper · 2026 · Wei Tian, Yuhao Zhou and Man Lan

Summary

The paper tackles two weaknesses of LLM-based Chinese Grammatical Error Correction (CGEC). General-purpose models know little about typical learner errors. Supervised fine-tuning with MLE also tends to rewrite sentences that need no change, which lowers precision. The authors propose CSRP, a three-stage pipeline built on Qwen3-4B:

  1. Continual pre-training on a deduplicated, filtered corpus mixing general Chinese text with error-correction data.
  2. Chain-of-Thought SFT, where Qwen-Plus supplies structured rationales that locate the error, classify it and explain it before the corrected sentence.
  3. GRPO reinforcement learning with an Efficiency-Aware Reward. Using Levenshtein distance, this reward credits edits that move the output closer to the reference and penalizes needless changes.

The resulting 4B model reaches state-of-the-art scores on NACGEC and CSCD, beating larger 14B systems and GPT-4. Its gains come mainly from higher precision, while recall stays almost unchanged. This matters because it shows that explicitly optimizing for minimal edits curbs over-correction more effectively than adding model size or training data. Code is released on GitHub (TW-NLP/ChineseErrorCorrector).

Key points

  • Pre-training data: the CPT corpus was cut from about 7.29M to 5,901,700 samples, mixing general and correction data at 8:2. Sources are wiki-zh-25, cci2 and lang8+HSK; wiki-zh-23 was dropped as too redundant.
  • Rationale quality: in a human check of 1,000 rationales, 95.2% were judged linguistically faithful (Cohen's κ = 0.81).
  • Main results: NACGEC reaches 50.99 F0.5 with 57.17 precision, compared with a prior plateau of roughly 45–46 F0.5. CSCD reaches 59.61 F1, 5.20 points above GPT-4.
  • Effect of GRPO: precision rose by +8.44 on NACGEC and +7.37 on CSCD, while recall fell by only 0.20 and 0.72. Applying GRPO without CPT gives just 45.97 F0.5, so the two stages contribute independently.
  • Cost trade-off: sampling 4 candidates instead of 8 during RL halves sampling cost and lowers F0.5 only slightly, to 50.61.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓East China Normal UniversityShanghai · China50 works · 2,174 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Wei Tian, Yuhao Zhou and Man Lan. 2026. CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1733

@inproceedings{tian2026csrp,
  title = {{CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards}},
  author = {Wei Tian, Yuhao Zhou and Man Lan},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1733},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1733}
}

References

38 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  2. Yayu Cao, Tianxiang Wang, Lvxiaowei Xu, Zhenyao Wang, and Ming Cai. 2025. Cxggec: Construction-guided grammatical error correction. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6143–6156. find
  3. Ming Dong, Zhiwei Cheng, Changyin Luo, and Tingting He. 2025. Retrieval-augmented generation for large language model based few-shot chinese spell checking. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10767– 10780. find
  4. Meiyuan Fang, Kai Fu, Jiping Wang, Yang Liu, Jin Huang, and Yitao Duan. 2020. A hybrid system for nlptea-2020 cged shared task. In Proceedings of the 6th Workshop on Natural Language Processing Techniques for Educational Applications, pages 67– 77. find
  5. Yong Hu, Fandong Meng, and Jie Zhou. 2022. Cscd-ime: correcting spelling errors generated by pinyin ime. arXiv preprint arXiv:2211.08788. — [arXiv:2211.08788](https://arxiv.org/abs/2211.08788) find
  6. Li Huang, Junjie Li, Weiwei Jiang, Zhiyu Zhang, Minchuan Chen, Shaojun Wang, and Jing Xiao. 2021. Phmospell: Phonological and morphological knowledge guided chinese spelling check. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International J… find
  7. Tuo Ji, Hang Yan, and Xipeng Qiu. 2021. Spellbert: A lightweight pretrained model for chinese spelling check. In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 3544–3551. find
  8. Zixiao Kong, Xianquan Wang, Shuanghong Shen, Keyu Zhu, Huibo Xu, and Yu Su. 2025. Scholargec: Enhancing controllability of large language model for chinese academic grammatical error correction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24339–24347. find
  9. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th annual meeting of the…
  10. Jing Li, Gaosheng Wu, Dafei Yin, Haozhao Wang, and Yonggang Wang. 2021. Dcspell: A detector-corrector framework for chinese spelling error correction. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1870–1874. find
  11. Kunting Li, Yong Hu, Liang He, Fandong Meng, and Jie Zhou. 2024. C-llm: learn to check chinese spelling errors character by character. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5944–5957. find
  12. Wei Li, Wen Luo, Guangyue Peng, and Houfeng Wang. 2025. Explanation based in-context demonstrations retrieval for multilingual grammatical error correction. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language T… find
  13. Wei Li and Houfeng Wang. 2024. Detection-correction structure via general language model for grammatical error correction. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1748–1763. find
  14. Junhong Liang and Yu Zhou. 2025. Rair: Retrieval-augmented iterative refinement for chinese spelling correction. arXiv preprint arXiv:2504.18938. — [arXiv:2504.18938](https://arxiv.org/abs/2504.18938) find
  15. Zihong Liang, Xiaojun Quan, and Qifan Wang. 2023. Disentangled phonetic representation for chinese spelling correction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13509– 13521. find
  16. Linfeng Liu, Hongqiu Wu, and Hai Zhao. 2024. Chinese spelling correction as rephrasing language model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18662–18670. find
  17. Shulin Liu, Tao Yang, Tianchi Yue, Feng Zhang, and Di Wang. 2021. Plome: Pre-training with misspelled knowledge for chinese spelling correction. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language… find
  18. Shirong Ma, Yinghui Li, Rongyi Sun, Qingyu Zhou, Shulin Huang, Ding Zhang, Li Yangning, Ruiyang Liu, Zhongli Li, Yunbo Cao, and 1 others. 2022. Linguistic rules-based corpus generation for native chinese grammatical error correction. In Findings of the Association for Computational Linguistics: EMNL…
  19. Ziheng Qiao, Houquan Zhou, and Zhenghua Li. 2025. Mixture of small and large models for chinese spelling check. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 28298– 28311. find
  20. Yixiao Song, Kalpesh Krishna, Rajesh Bhatt, Kevin Gimpel, and Mohit Iyyer. 2024. Gee! grammar error explanation with large language models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 754–781. find
  21. Chang Su, Xiaofeng Zhao, Xiaosong Qiao, Min Zhang, Hao Yang, Junhao Zhu, Ming Zhu, and Wenbing Ma. 2023. Hwcgec: hw-tsc’s 2023 submission for the nlpcc2023’s chinese grammatical error correction task. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 59–68.… find
  22. Min Tan, Dagang Chen, Zesong Li, and Peng Wang. 2020. Spelling error correction with bert based on character-phonetic. In 2020 IEEE 6th International Conference on Computer and Communications (ICCC), pages 1146–1150. IEEE. find
  23. Wei Tian and 1 others. 2025. Chineseerrorcorrector3-4b: State-of-the-art chinese spelling and grammar corrector. arXiv preprint arXiv:2511.17562. — [arXiv:2511.17562](https://arxiv.org/abs/2511.17562) find
  24. Yuen-Hsien Tseng, Lung-Hao Lee, Li-Ping Chang, and Hsin-Hsi Chen. 2015. Introduction to sighan 2015 bake-off for chinese spelling check. In Proceedings of the Eighth SIGHAN Workshop on Chinese Language Processing, pages 32–37. find
  25. Dingmin Wang, Yan Song, Jing Li, Jialong Han, and Haisong Zhang. 2018. A hybrid approach to automatic corpus generation for chinese spelling check. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2517–2527.
  26. Chi Wei, Shaobin Huang, Rongsheng Li, Naiyu Yan, and Rui Wang. 2024. Training a better chinese spelling correction model via prior-knowledge guided teacher. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13578– 13589. find
  27. Xiao Wei, Jianbao Huang, Hang Yu, and Qian Liu. 2023. Ptcspell: Pre-trained corrector based on character shape and pinyin for chinese spelling correction. In Findings of the Association for Computational Linguistics: ACL 2023, pages 6330–6343. find
  28. Yue Zhang, Bo Zhang, Zhenghua Li, Zuyi Bao, Chen Li, and Min Zhang. 2022. Syngec: Syntax-enhanced grammatical error correction with a tailored gec-oriented parser. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2518–2531.
  29. Yuanyuan Zhao, Nan Jiang, Weiwei Sun, and Xiaojun Wan. 2018. Overview of the nlpcc 2018 shared task: Grammatical error correction. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 439–445. Springer. find
  30. Houquan Zhou, Bo Zhang, Zhenghua Li, Ming Yan, and Min Zhang. 2025. A training-free llm-based approach to general chinese character error correction. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13827–13852. find
  31. Cheng Wen, Xianghui Sun, Shuaijiang Zhao, Xiaoquan Fang, Liangyu Chen, and Wei Zou. 2023. Chathome: Development and evaluation of a domain-specific language model for home renovation. arXiv preprint arXiv:2307.15290. — [arXiv:2307.15290](https://arxiv.org/abs/2307.15290) find
  32. Heng-Da Xu, Zhongli Li, Qingyu Zhou, Chao Li, Zizhen Wang, Yunbo Cao, He-Yan Huang, and Xian-Ling Mao. 2021. Read, listen, and see: Leveraging multimodal information helps chinese spell checking. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 716–728. find
  33. Lvxiaowei Xu, Jianwang Wu, Jiawei Peng, Jiayu Fu, and Ming Cai. 2022. Fcgec: Fine-grained corpus for chinese grammatical error correction. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1900–1918. find
  34. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  35. Baolin Zhang. 2009. Features and functions of the hsk dynamic composition corpus. International Chinese Language Education, 4:71–79. find
  36. Shaohua Zhang, Haoran Huang, Jicong Liu, and Hang Li. 2020. Spelling error correction with soft-masked bert. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 882–890.
  37. Yue Zhang, Bo Zhang, Haochen Jiang, Zhenghua Li, Chen Li, Fei Huang, and Min Zhang. 2023. Nasgec: a multi-domain chinese grammatical error correction dataset from native speaker texts. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9935–9951. find
  38. A Dataset Train Test Type CSCD-NS 30,000 5,000 CSC WK271+SIGHAN15 251,835 1,100 CSC CGED 20,449 - CGEC NACGEC 7,568 5,869 CGEC FCGEC 35,355 - CGEC Subtotal 345,207 11,969 -After Deduplication 338,530 11,969 -After CoT Augmentation 336,604 11,946 -Final Split: SFT (80%) = 269,283; RL (20%) = 67,321 D… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tian-2026-csrp-chain-thought-reasoning-chinese/