| # CL²GEC:A Multi-Discipline Benchmark for Continual Learning in Chinese Literature Grammatical Error Correction | |
| **CL²GEC** is a benchmark for **Chinese grammatical error correction (GEC)** in **scholarly writing** with a **continual-learning** protocol. The corpus covers **10 first-level disciplines** (Law, Management, Education, Economics, Natural Sciences, History, Agricultural Sciences, Literature, Arts, Philosophy). Each sample contains an errorful sentence (`source`) and one or more corrected references (`references`). Standard **train / validation / test** splits are provided and may be used **per-discipline** to study sequential/continual learning behavior such as forgetting and transfer. | |
| --- | |
| ## Supported Tasks and Leaderboards | |
| **Grammatical Error Correction (GEC)** / **Text-to-Text Generation** | |
| - **Input**: a Chinese sentence containing grammatical/usage errors. | |
| - **Output**: a semantically equivalent, grammatically correct sentence. | |
| **Recommended Metrics** | |
| - GEC metrics: **Precision / Recall / F0.5** (e.g., via ChERRANT). | |
| - Continual-learning (optional): **Average Performance** and **Backward Transfer (BWT)** computed over task sequences defined by the ordered disciplines. | |
| --- | |
| ## Dataset Structure | |
| ### Data Instances | |
| Below is a recommended public JSON schema: | |
| ```json | |
| { | |
| "id": "0", | |
| "source": "总体上看，仍有许多案件以不适用调解制度。", | |
| "references": [ | |
| "总体上看，依然有许多案件不适宜使用调解制度来解决。" | |
| ], | |
| "category": "法学", | |
| "edits": [ | |
| { | |
| "src_interval": [7, 9], | |
| "tgt_interval": [7, 9], | |
| "src_content": ["不", "适", "用"], | |
| "tgt_content": ["不", "适", "宜"] | |
| } | |
| ] | |
| } | |
| ``` | |
| ### Data Fields | |
| - **id** *(string)*: unique sample identifier. | |
| - **source** *(string)*: original sentence with errors. | |
| - **references** *(list[string])*: one or more corrected sentences. | |
| - **category** *(string)*: first-level discipline. | |
| - **edits** *(list[object], optional)*: token/character-level edits (if provided). | |
| ### Data Splits | |
| | Split | #Samples | Notes | | |
| | ---------- | -------: | ------------------- | | |
| | train | 7,000 | training data | | |
| | validation | 1,000 | development set | | |
| | test | 2,000 | held-out evaluation | | |
| --- | |
| ## Categories (Disciplines) | |
| Below are the 10 discipline labels (Chinese) with suggested English names: | |
| | Chinese (label in data) | English | | |
| | ----------------------- | ---------- | | |
| | 法学 | Law | | |
| | 管理 | Management | | |
| | 教育 | Education | | |
| | 经济学 | Economics | | |
| | 理学 | Sciences | | |
| | 历史学 | History | | |
| | 农学 | Agronomy | | |
| | 文学 | Literature | | |
| | 哲学 | Philosophy | | |
| | 艺术学 | Arts | | |
| --- | |
| ## Collection and Annotation | |
| - **Sources**: Extracted from CNKI Academic PDFs, covering 10 first-level disciplines and 100 second-level disciplines; only abstracts and main text are retained; non-linguistic content such as references, acknowledgments, formulas, tables, and figure captions are removed; sentence-level segmentation uses LTP. Anonymization is also performed. | |
| - **Annotation**: | |
| 1. Multi-model consistency error detection to screen candidates (e.g., GECToR, Chinese-BART, etc.); | |
| 2. LLM pre-rewrite as weak references; | |
| 3. Dual independent annotation (by senior annotators with the same subject background), unifying style, revision, and merging; | |
| 4. 100% review by domain experts to ensure publication-level quality, supplementing with multiple references when necessary. | |
| --- | |
| ## Intended Uses | |
| - Research on **Chinese GEC** for scholarly prose. | |
| - Cross-domain robustness and **discipline-aware** modeling. | |
| - **Continual learning** studies focusing on forgetting/transfer across disciplines. | |
| --- | |
| ## Ethical Considerations & Privacy | |
| - Texts are anonymized and cleaned to remove sensitive information. | |
| - Sentences are taken from academic texts and contain academic terminology; when the model is made available for public use, the risks and scope of application should be declared and misuse should be avoided. | |
| - Ensure that upstream content complies with platform/journal usage policies and your chosen **license** clearly states permitted uses. | |
| --- | |
| ## Citation | |
| If you use this dataset in your research, please cite (replace with your paper details): | |
| ```bibtex | |
| @misc{qin2025cl2gec, | |
| title = {CL$^2$GEC: A Multi-Discipline Benchmark for Continual Learning in Chinese Literature Grammatical Error Correction}, | |
| author = {Shang Qin and Jingheng Ye and Yinghui Li and Hai-Tao Zheng and Qi Li and Jinxiao Shan and Zhixing Li and Hong-Gee Kim}, | |
| year = {2025}, | |
| eprint = {2509.13672}, | |
| archivePrefix = {arXiv}, | |
| primaryClass = {cs.CL}, | |
| url = {https://arxiv.org/abs/2509.13672} | |
| } | |
| ``` | |
| --- | |
| ## Changelog | |
| - **v1.0.0**: initial public release; includes train/validation/test splits, field schema, usage examples, and evaluation guidance. |