# 中文拼写和语法纠错 [**🇨🇳中文**](https://github.com/TW-NLP/ChineseErrorCorrector/blob/main/README.md)
Logo
----------------- ## 介绍 支持中文拼写和语法错误纠正，并开源拼写和语法错误的增强工具、大模型训练代码。荣获2024CCL 冠军 🏆，[查看论文](https://aclanthology.org/2024.ccl-3.31/) ，[2023 NLPCC-NaCGEC纠错冠军🏆](https://github.com/TW-NLP/ChineseErrorCorrector?tab=readme-ov-file#nacgec-%E6%95%B0%E6%8D%AE%E9%9B%86)， [2022 FCGEC 纠错冠军🏆](https://github.com/TW-NLP/ChineseErrorCorrector?tab=readme-ov-file#fcgec-%E6%95%B0%E6%8D%AE%E9%9B%86) ，如有帮助，感谢star✨。 ## 🔥🔥🔥 新闻 [2025/04/28] 根据[建议](https://github.com/TW-NLP/ChineseErrorCorrector/issues/17) ，我们重新训练纠错模型，并完全开源训练步骤，支持结果复现，[复现教程](https://github.com/TW-NLP/ChineseErrorCorrector/tree/main?tab=readme-ov-file#%E5%AE%9E%E9%AA%8C%E7%BB%93%E6%9E%9C%E5%A4%8D%E7%8E%B0) [2025/03/17] 更新批量错误文本的解析，[transformers批量解析](https://github.com/TW-NLP/ChineseErrorCorrector?tab=readme-ov-file#transformers-%E6%89%B9%E9%87%8F%E6%8E%A8%E7%90%86) ;[VLLM批量解析](https://github.com/TW-NLP/ChineseErrorCorrector?tab=readme-ov-file#vllm-%E5%BC%82%E6%AD%A5%E6%89%B9%E9%87%8F%E6%8E%A8%E7%90%86) [2025/03/10] 模型支持多种推理方式，包括 transformers、VLLM、modelscope。 [2025/02/25] 🎉🎉🎉使用200万纠错数据进行多轮迭代训练，发布了[twnlp/ChineseErrorCorrector2-7B](https://huggingface.co/twnlp/ChineseErrorCorrector2-7B) ，在 [NaCGEC-2023NLPCC官方评测数据集](https://github.com/masr2000/NaCGEC) 上，超越第一名华为10个点，遥遥领先，推荐使用✨✨， [技术详情](https://blog.csdn.net/qq_43765734/article/details/145858955) [2025/02] 为方便部署，使用38万开源拼写数据，发布了[twnlp/ChineseErrorCorrector-1.5B](https://huggingface.co/twnlp/ChineseErrorCorrector-1.5B) [2025/01] 使用38万开源拼写数据，基于Qwen2.5训练中文拼写纠错模型，支持语似、形似等错误纠正，发布了[twnlp/ChineseErrorCorrector-7B](https://huggingface.co/twnlp/ChineseErrorCorrector-7B)，[twnlp/ChineseErrorCorrector-32B-LORA](https://huggingface.co/twnlp/ChineseErrorCorrector-32B-LORA/tree/main) [2024/06] v0.1.0版本：🎉🎉🎉开源一键语法错误增强工具，该工具可以进行14种语法错误的增强，不同行业可以根据自己的数据进行错误替换，来训练自己的语法和拼写模型。详见[Tag-v0.1.0](https://github.com/TW-NLP/ChineseErrorCorrector/tree/0.1.0) ## 模型列表 | 模型名称 | 纠错类型 | 描述 | |:--------------------------------------------------------------------------------------------|:------|:------------------------------------------| | [twnlp/ChineseErrorCorrector2-7B](https://huggingface.co/twnlp/ChineseErrorCorrector2-7B) | 语法+拼写 | 使用200万纠错数据进行多轮迭代训练，适用于语法纠错和拼写纠错，效果好，推荐使用。 | | [twnlp/ChineseErrorCorrector-7B](https://huggingface.co/twnlp/ChineseErrorCorrector-7B) | 拼写 | 使用38万开源拼写数据，支持语似、形似等拼写错误纠正，拼写纠错效果好。 | | [twnlp/ChineseErrorCorrector-1.5B](https://huggingface.co/twnlp/ChineseErrorCorrector-1.5B) | 拼写 | 使用38万开源拼写数据，支持语似、形似等拼写错误纠正，拼写纠错效果一般。 | ## 数据集 | 数据集名称 | 数据链接 | 数据量和类别说明 | 描述 | |:-----------------------------|:-------------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------|:--------------------------------| | ChinseseErrorCorrectData | [twnlp/ChinseseErrorCorrectData](https://huggingface.co/datasets/twnlp/ChinseseErrorCorrectData) | 200万 | ChineseErrorCorrector2-7B 训练数据集 | | CSC（拼写纠错数据集） | [twnlp/csc_data](https://huggingface.co/datasets/twnlp/csc_data) | W271K(279,816) Medical(39,303) Lemon(22,259) ECSpell(6,688) CSCD(35,001) | 中文拼写纠错的数据集 | | CGC（语法纠错数据集） | [twnlp/cgc_data](https://huggingface.co/datasets/twnlp/cgc_data) | CGED(20,449) FCGEC(37,354) MuCGEC(2,467) NaSGEC(7,568) | 中文语法纠错的数据集 | | Lang8+HSK（百万语料-拼写和语法错误混合数据集） | [twnlp/lang8_hsk](https://huggingface.co/datasets/twnlp/lang8_hsk) | 1,568,885 | 中文拼写和语法数据集 | ## 拼写纠错评测 - 评估指标：F1 | Model Name | Model Link | Base Model | Avg | SIGHAN-2015(通用) | EC-LAW(法律) | EC-MED(医疗) | EC-ODW(公文) | |:-------------------------------------|:-------------------------------------------------------------------------------------|:---------------------------|:------|:----------------|:-----------|:-----------|:-----------| | twnlp/ChineseErrorCorrector-1.5B | [huggingface](https://huggingface.co/twnlp/ChineseErrorCorrector-1.5B/tree/main) | Qwen/Qwen2.5-1.5B-Instruct | 0.459 | 0.346 | 0.517 | 0.433 | 0.540 | | twnlp/ChineseErrorCorrector-7B | [huggingface](https://huggingface.co/twnlp/ChineseErrorCorrector-7B/tree/main) | Qwen/Qwen2.5-7B-Instruct | 0.712 | 0.592 | 0.787 | 0.677 | 0.793 | | twnlp/ChineseErrorCorrector-32B-LORA | [huggingface](https://huggingface.co/twnlp/ChineseErrorCorrector-32B-LORA/tree/main) | Qwen/Qwen2.5-32B-Instruct | 0.757 | 0.594 | 0.776 | 0.794 | 0.864 | ## 文本纠错评测(双冠军 🏆) ### NaCGEC 数据集 - 评估工具：ChERRANT [评测工具](https://github.com/HillZhang1999/MuCGEC) - 评估数据：[NaCGEC](https://github.com/masr2000/NaCGEC) - 评估指标：F1-0.5 🏆 | Model Name | Model Link | Prec | Rec | F0.5 | |:-----------------|:---------------------------------------------------------------|:-----------|:------------|:-------| | twnlp/ChineseErrorCorrector2-7B | [huggingface](https://huggingface.co/twnlp/ChineseErrorCorrector2-7B) ； [modelspose(国内下载)](https://www.modelscope.cn/models/tiannlp/ChineseErrorCorrector2-7B) | 0.5686 | 0.57 | 0.5689 | | HW_TSC_nlpcc2023_cgec(华为) | 未开源 | 0.5095 | 0.3129 | 0.4526 | | 鱼饼啾啾Plus(北京大学) | 未开源 | 0.5708 | 0.1294 | 0.3394 | | CUHK_SU(香港中文大学) | 未开源 | 0.3882 | 0.1558 | 0.2990 | ### FCGEC 数据集 - 评估指标：binary_f1 [评测🏆](https://codalab.lisn.upsaclay.fr/competitions/8020#results) ## 使用 ### 🤗 transformers ```shell pip install transformers ``` ```shell from transformers import AutoModelForCausalLM, AutoTokenizer,set_seed set_seed(42) model_name = "twnlp/ChineseErrorCorrector2-7B" model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="auto", device_map="auto" ) tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side='left') prompt = "你是一个文本纠错专家，纠正输入句子中的语法错误，并输出正确的句子，输入句子为：" text_input = "对待每一项工作都要一丝不够。" messages = [ {"role": "user", "content": prompt + text_input} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) model_inputs = tokenizer([text], return_tensors="pt").to(model.device) generated_ids = model.generate( **model_inputs, max_new_tokens=512 ) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) ] response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] print(response) ``` ### VLLM ```shell pip install transformers pip install vllm==0.3.3 ``` ```shell from transformers import AutoTokenizer from vllm import LLM, SamplingParams # Initialize the tokenizer tokenizer = AutoTokenizer.from_pretrained("twnlp/ChineseErrorCorrector2-7B") # Pass the default decoding hyperparameters of twnlp/ChineseErrorCorrector2-7B # max_tokens is for the maximum length for generation. sampling_params = SamplingParams(seed=42,max_tokens=512) # Input the model name or path. Can be GPTQ or AWQ models. llm = LLM(model="twnlp/ChineseErrorCorrector2-7B") # Prepare your prompts text_input = "对待每一项工作都要一丝不够。" messages = [ {"role": "user", "content": "你是一个文本纠错专家，纠正输入句子中的语法错误，并输出正确的句子，输入句子为："+text_input} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) # generate outputs outputs = llm.generate([text], sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` ### VLLM 异步批量推理 - Clone the repo ``` sh git clone https://github.com/TW-NLP/ChineseErrorCorrector cd ChineseErrorCorrector ``` - Install Conda: please see https://docs.conda.io/en/latest/miniconda.html - Create Conda env: ``` sh conda create -n zh_correct -y python=3.10 conda activate zh_correct pip install -r requirements.txt # If you are in mainland China, you can set the mirror as follows: pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com ``` ```sh # 修改config.py #（1）根据不同的模型，修改的DEFAULT_CKPT_PATH，默认为ChineseErrorCorrector2-7B(将模型下载，放在ChineseErrorCorrector/pre_model/ChineseErrorCorrector2-7B) #（2）将Qwen2TextCorConfig的USE_VLLM = True #批量预测 python main.py ``` ### Transformers 批量推理 - Clone the repo ``` sh git clone https://github.com/TW-NLP/ChineseErrorCorrector cd ChineseErrorCorrector ``` - Install Conda: please see https://docs.conda.io/en/latest/miniconda.html - Create Conda env: ``` sh conda create -n zh_correct -y python=3.10 conda activate zh_correct pip install -r requirements.txt # If you are in mainland China, you can set the mirror as follows: pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com ``` ``` sh # 修改config.py #（1）根据不同的模型，修改的DEFAULT_CKPT_PATH，默认为ChineseErrorCorrector2-7B #（2）将Qwen2TextCorConfig的USE_VLLM = False #批量预测 python main.py #输出： ''' [{'source': '对待每一项工作都要一丝不够。', 'target': '对待每一项工作都要一丝不苟。', 'errors': [('够', '苟', 12)]}, {'source': '大约半个小时左右', 'target': '大约半个小时', 'errors': [('左右', '', 6)]}] ''' ``` ### 🤖 modelscope ```shell pip install modelscope ``` ```shell from modelscope import AutoModelForCausalLM, AutoTokenizer model_name = "tiannlp/ChineseErrorCorrector2-7B" model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="auto", device_map="auto" ) tokenizer = AutoTokenizer.from_pretrained(model_name) prompt = "你是一个文本纠错专家，纠正输入句子中的语法错误，并输出正确的句子，输入句子为：" text_input = "对待每一项工作都要一丝不够。" messages = [ {"role": "user", "content": prompt + text_input} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) model_inputs = tokenizer([text], return_tensors="pt").to(model.device) generated_ids = model.generate( **model_inputs, max_new_tokens=512 ) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) ] response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] print(response) ``` ## 实验结果复现 ### 环境准备 - Clone the repo ``` sh git clone https://github.com/TW-NLP/ChineseErrorCorrector cd ChineseErrorCorrector ``` - Install Conda: please see https://docs.conda.io/en/latest/miniconda.html - Create Conda env: ``` sh conda create -n zh_correct -y python=3.10 conda activate zh_correct pip install -r requirements.txt # If you are in mainland China, you can set the mirror as follows: pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com ``` ### 数据和模型的准备 1、下载训练数据集：[twnlp/ChinseseErrorCorrectData](https://huggingface.co/datasets/twnlp/ChinseseErrorCorrectData) ,放在 `/data/paper_data` 中。 2、下载Qwen2.5-7B-Instruct：[Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) ,放在`/pre_model`中 ### 模型训练与合并 ``` sh # Lang8+HSK 训练 bash ./llm/train/run1.sh bash ./llm/train/merge1.sh # CGC+CSC 数据集训练 bash ./llm/train/run2.sh bash ./llm/train/merge2.sh # Nacgec 数据集训练 bash ./llm/train/run3.sh bash ./llm/train/merge3.sh ``` ## Citation If this work is helpful, please kindly cite as: ```bibtex @inproceedings{wei2024中小学作文语法错误检测, title={中小学作文语法错误检测, 病句改写与流畅性评级的自动化方法研究}, author={Wei, Tian}, booktitle={Proceedings of the 23rd Chinese National Conference on Computational Linguistics (Volume 3: Evaluations)}, pages={278--284}, year={2024} } ``` ## Star History ![Star History Chart](https://api.star-history.com/svg?repos=TW-NLP/ChineseErrorCorrector&type=Date)