一键让中文大模型化身中文拼写纠错模型!!!
本仓库提供了以下论文的实现：
- (v2.0.0) A Training-free LLM-based Approach to General Chinese Character Error Correction.
- (v1.0.0) A Simple yet Effective Training-free Prompt-free Approach to Chinese Spelling Correction Based on Large Language Models.
新闻
- 2025/02/21: 新版本 (v2.0.0) 发布。
- 现在我们支持字符的插入和删除操作。
- 添加了 Prompt-based 模型概率，提高纠错的性能。
- 2024/12/09: 我们荣获金山办公2024算法挑战赛-中文文本智能校对大赛冠军。在比赛中，我们以本代码库作为核心模块，取得了
$F_{0.5}$ 分数比第二名高 2.02 分的成绩。 - 2024/09/20: 我们的论文被 EMNLP 2024 主会 接收。
- torch>=2.0.1
- transformers>=4.27.0
- xformers==0.0.21
- accelerate
- bitsandbytes
- sentencepiece
- pypinyin
- pypinyin-dict
- opencc-python-reimplemented
- modelscope (可选，用于从 modelscope 下载模型)
- streamlit (可选，用于演示应用)
- uvicorn (可选，用于 RESTful API 服务器)
- fastapi (可选，用于 RESTful API 服务器)
- loguru (可选，用于 RESTful API 服务器)
- sse_starlette (可选，用于 RESTful API 服务器)
您可以通过运行以下命令来配置环境：
bash scripts/set_environment.sh这将自动创建虚拟环境并安装所需的包。
为获得更好的性能，您可以安装 flash-attn：
pip install flash-attn --no-build-isolationWarning
一位用户报告说，使用 Qwen2 或 Qwen2.5 系列模型时，如果没有安装 flash-attn，代码会产生意料之外的行为。
请安装 flash-attn 来避免这个问题。或者您可以在 LMCorrector 类中设置 torch_dtype=torch.bfloat16 来避免这个问题。
虽然我们强烈建议使用 flash-attn，它将显著减少显存使用并加快推理速度。
我们为代码打上了两个标签：
- v1.0.0: 第一个版本，对应论文 A Simple yet Effective Training-free Prompt-free Approach to Chinese Spelling Correction Based on Large Language Models.
- v2.0.0: 第二个版本，对应论文 A Training-free LLM-based Approach to General Chinese Character Error Correction.
代码库将不断更新。 如果您想使用特定版本，可以使用以下命令切换到特定版本：
git checkout v1.0.0或者
git checkout v2.0.0如果在本地缓存中未找到模型，代码将自动从 Huggingface 模型仓库下载模型。
Tip
在我们的实验中，我们发现 Base 版本的 LLM 比 Instruct 或 Chat 版本的 LLM 表现更好。因此我们推荐使用 Base 版本的模型。
推荐模型:
v1.0.0: 该版本中使用baichuan-inc/Baichuan2-13B-Base可以获得最好的性能。v2.0.0: 使用Qwen/Qwen2.5-[7/14]B可以在性能和速度之间取得一个很好的平衡。
我们提供了一个简单的 Python API 用于纠错：
from lmcsc import LMCorrector
corrector = LMCorrector(
model="Qwen/Qwen2.5-1.5B",
prompted_model="Qwen/Qwen2.5-1.5B", # 建议 model 和 prompted_model 使用相同的模型以减少显存占用。
config_path="configs/c2ec_config.yaml", # 你可以修改为 `default_config.yaml` 来禁用添字、删字操作。
torch_dtype=torch.bfloat16, # 默认我们会使用 torch.float16 来进行计算。 但是我们发现在使用 Qwen2 或者 Qwen2.5 在没安装 flash-attn 的时候会导致运行时间过长。如果已经安装了 flash-attn 你可以不用专门设置这个参数。
)
outputs = corrector("完善农产品上行发展机智。")
print(outputs)
# [('完善农产品上行发展机制。',)]也支持流式模式：
outputs = corrector("完善农产品上行发展机智。", stream=True)
for output in outputs:
print(output[0][0], end="\r", flush=True)
print()我们还提供了纠错器的 RESTful API 服务器。
# 建议 model 和 prompted_model 使用相同的模型以减少显存占用。
# 你可以修改为 `default_config.yaml` 来禁用添字、删字操作。
# 使用 bfloat16 来避免 Qwen2 或 Qwen2.5 系列模型在未安装 flash-attn 时产生意料之外的行为。
python api_server.py \
--model "Qwen/Qwen2.5-1.5B" \
--prompted_model "Qwen/Qwen2.5-1.5B" \
--config_path "configs/c2ec_config.yaml" \
--host 127.0.0.1 \
--port 8000 \
--workers 1 \
--bf16您可以使用 curl 来测试 RESTful API 服务器。
# 默认模式
curl -X POST 'http://127.0.0.1:8000/correction' -H 'Content-Type: application/json' -d '{"input": "完善农产品上行发展机智。"}'
# > {"id":"","object":"correction","choices":[{"index":0,"message":{"content":"完善农产品上行发展机制。"}}],"created":1727058762}
# 流式模式
curl -X POST 'http://127.0.0.1:8000/correction' -H 'Content-Type: application/json' -d '{"input": "完善农产品上行发展机智。", "stream": "True"}'
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展机制"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展模式。"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展机制。"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展机制。"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展机制。"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"完善农产品上行发展机制。"},"index":0}],"created":1727058762}
# > data: [DONE]
# 带上下文的纠错
curl -X POST 'http://127.0.0.1:8000/correction' -H 'Content-Type: application/json' -d '{"input": "未挨前兆", "contexts": "患者提问：", "stream": "True"}'
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"未"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"未挨"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"胃癌前"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"胃癌前兆"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"胃癌前兆"},"index":0}],"created":1727058762}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"胃癌前兆"},"index":0}],"created":1727058763}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"胃癌前兆"},"index":0}],"created":1727058763}
# > data: {"id":"","object":"correction.chunk","choices":[{"delta":{"content":"胃癌前兆"},"index":0}],"created":1727058763}
# > data: [DONE]我们为我们的方法提供了一个演示应用 (Demo)。要运行该 Demo，您需要：
- 确保您已安装
streamlit包。 - 运行以下命令：
streamlit run demo.py默认情况下，Demo 所使用 Qwen/Qwen2.5-0.5B，可以在具有 32GB 内存的 V100 GPU 上运行。您可以在演示的侧边栏中或通过修改 configs/demo_app_config.yaml 中的 default_model 来更换其他模型。
Demo 的纠错配置为 configs/demo_config.yaml，这与 v2.0.0 中的 c2ec_config.yaml 相同。
您还可以修改 configs/demo_app_config.yaml 中的 config_path 来使用其他配置。
侧边栏还允许您调整 n_beam、alpha 和 use_faithfulness_reward 参数。
侧边栏中提供了几个示例，包括一个包含 1866 个字的长句。
可以通过以下命令运行论文中提到的数据集的实验：
python -u run.py \
--input-file <input-file> \
--path <path> \
--model-name <model-name> \
--prompted-model-name <prompted-model-name> \
--config-path <config-path> \
--n-observed-chars <n-observed-chars> \
--n-beam <n-beam> \
--batch-size <batch-size> \
--alpha <alpha> \
--use-faithfulness-reward运行之前，您需要将每个句子对预处理成以下格式：
[src] [tgt]
[src] [tgt]
[src] [tgt]
其中 [src] 和 [tgt] 分别是源句子和目标句子。
使用 \t 分隔它们。
数据准备过程可以在 scripts/download_datasets.sh 中找到。
该脚本将从原始来源（托管在 raw.githubusercontent.com 和 Google Drive 上）下载数据集，并将它们预处理成所需格式。
我们构造了一个 C2EC 数据来测试我们方法处理多字漏字的效果。 你可以通过如下步骤进行构建。
-
从 CTCResource 下载 CCTC v1.1 数据集。
- 将下载下来的
cctc_v1.1.zip(MD5:ecd94ad85c33d7c0ace11b6da316f81e) 文件放入data/c2ec文件夹。你不需要去解压它，后续程序会帮你解压。 - 在放入 cctc_v1.1.zip 文件后，
data/c2ec的目录结构应该看起来长这样:datasets/ c2ec/ metadata/ dev.index test.index cctc_v1.1.zip
- 将下载下来的
-
运行下面命令来构造构造数据集。在运行过程中 Lemon 会被自动从 Lemon 中下载。
bash scripts/build_c2ec_dataset.sh
-
最后你就得到了 C2EC 数据集了。
Note
我们利用 CCTC 和 Lemon 构造了 C2EC 数据集。 如果您在文字中使用了 C2EC 数据也请引用他们的原始论文 :D
-
CCTC Dataset
@inproceedings{wang-etal-2022-cctc, title = "{CCTC}: A Cross-Sentence {C}hinese Text Correction Dataset for Native Speakers", author = "Wang, Baoxin and Duan, Xingyi and Wu, Dayong and Che, Wanxiang and Chen, Zhigang and Hu, Guoping", editor = "Calzolari, Nicoletta and Huang, Chu-Ren and Kim, Hansaem and Pustejovsky, James and Wanner, Leo and Choi, Key-Sun and Ryu, Pum-Mo and Chen, Hsin-Hsi and Donatelli, Lucia and Ji, Heng and Kurohashi, Sadao and Paggio, Patrizia and Xue, Nianwen and Kim, Seokhwan and Hahm, Younggyun and He, Zhong and Lee, Tony Kyungil and Santus, Enrico and Bond, Francis and Na, Seung-Hoon", booktitle = "Proceedings of the 29th International Conference on Computational Linguistics", month = oct, year = "2022", address = "Gyeongju, Republic of Korea", publisher = "International Committee on Computational Linguistics", url = "https://aclanthology.org/2022.coling-1.294/", pages = "3331--3341" }
-
Lemon Dataset
@inproceedings{wu-etal-2023-rethinking, title = "Rethinking Masked Language Modeling for {C}hinese Spelling Correction", author = "Wu, Hongqiu and Zhang, Shaohua and Zhang, Yuchen and Zhao, Hai", editor = "Rogers, Anna and Boyd-Graber, Jordan and Okazaki, Naoaki", booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)", month = jul, year = "2023", address = "Toronto, Canada", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2023.acl-long.600/", doi = "10.18653/v1/2023.acl-long.600", pages = "10743--10756" }
- GPT2
- Baichuan2
- Qwen1.5
- Qwen2
- Qwen2.5
- InternLM2
- 允许字的插入和删除。
- 加快推理过程。
- 将代码打包成库。
- 允许 continuous batching。
欢迎贡献！请随时提交 Pull Request。
该项目采用 Apache 许可证 - 详见 LICENSE 文件。
如果您觉得这个工作有用，请引用：
@inproceedings{zhou-etal-2024-simple,
title = "A Simple yet Effective Training-free Prompt-free Approach to {C}hinese Spelling Correction Based on Large Language Models",
author = "Zhou, Houquan and
Li, Zhenghua and
Zhang, Bo and
Li, Chen and
Lai, Shaopeng and
Zhang, Ji and
Huang, Fei and
Zhang, Min",
editor = "Al-Onaizan, Yaser and
Bansal, Mohit and
Chen, Yun-Nung",
booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2024",
address = "Miami, Florida, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.emnlp-main.966",
pages = "17446--17467",
abstract = "This work proposes a simple training-free prompt-free approach to leverage large language models (LLMs) for the Chinese spelling correction (CSC) task, which is totally different from all previous CSC approaches. The key idea is to use an LLM as a pure language model in a conventional manner. The LLM goes through the input sentence from the beginning, and at each inference step, produces a distribution over its vocabulary for deciding the next token, given a partial sentence. To ensure that the output sentence remains faithful to the input sentence, we design a minimal distortion model that utilizes pronunciation or shape similarities between the original and replaced characters. Furthermore, we propose two useful reward strategies to address practical challenges specific to the CSC task. Experiments on five public datasets demonstrate that our approach significantly improves LLM performance, enabling them to compete with state-of-the-art domain-general CSC models.",
}