README.md

---
license: afl-3.0
language:
- zh
tags:
- Chinese Spell Correction
- csc
- Chinese Spell Checking
---

# ReaLiSe-for-csc

中文拼写纠错(Chinese Spell Checking, CSC)模型

该模型源于ReaLiSe源码提供的模型

原论文为：https://arxiv.org/abs/2105.12306

原论文官方代码为：https://github.com/DaDaMrX/ReaLiSe

本模型在SIGHAN2015上的表现如下：

|  | Detect-Acc | Detect-Precision | Detect-Recall | Detect-F1 | Correct-Acc | Correct-Precision | Correct-Recall | Correct-F1 | 
|--|--|--|--|--|--|--|--|--|
| Sentence-level | 84.7 | 77.3 | 81.3 | 79.3 | 84.0 | 75.9 | 79.9 | 77.8 |



# 模型使用方法

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/iioSnail/SCOPE/blob/main/ChineseBERT-for-csc_Demo.ipynb)


安装依赖：

```
!pip install transformers
!pip install pypinyin
!pip install boto3
```

```
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("iioSnail/ReaLiSe-for-csc", trust_remote_code=True)
model = AutoModel.from_pretrained("iioSnail/ReaLiSe-for-csc", trust_remote_code=True)

inputs = tokenizer(["我是炼习时长两念半的个人练习生蔡徐坤"], return_tensors='pt')
output_hidden = model(**inputs).logits
print(''.join(tokenizer.convert_ids_to_tokens(output_hidden.argmax(-1)[0, 1:-1])))
```

输出：

```
我是练习时长两年半的个人练习生蔡徐坤
```

你也可以使用本模型封装的`predict`方法。

```
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("iioSnail/ReaLiSe-for-csc", trust_remote_code=True)
model = AutoModel.from_pretrained("iioSnail/ReaLiSe-for-csc", trust_remote_code=True)

model.set_tokenizer(tokenizer)  # 使用predict方法前，调用该方法
print(model.predict("我是练习时长两念半的鸽仁练习生蔡徐坤"))
print(model.predict(["我是练习时长两念半的鸽仁练习生蔡徐坤", "喜换唱跳、rap 和 蓝球"]))
```

输出：

```
我是练习时长两年半的各仁练习生蔡徐坤
['我是练习时长两年半的各仁练习生蔡徐坤', '喜欢唱跳、rap 和 蓝球']
```

# 模型训练

```
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("iioSnail/ReaLiSe-for-csc", trust_remote_code=True)
model = AutoModel.from_pretrained("iioSnail/ReaLiSe-for-csc", trust_remote_code=True)

inputs = tokenizer(["我是炼习时长两念半的个人练习生蔡徐坤", "喜换唱跳rap蓝球"],
                   text_target=["我是练习时长两年半的个人练习生蔡徐坤", "喜欢唱跳rap篮球"],
                   padding=True,
                   return_tensors='pt')
loss = model(**inputs).loss
print("loss:", loss)
loss.backward()
```

输出：

```
loss: tensor(0.6515, grad_fn=<NllLossBackward0>)
```


# 常见问题

1. 网络问题，例如：`Connection Error`

解决方案：将模型下载到本地使用。批量下载方案可参考该[博客](https://blog.csdn.net/zhaohongfei_358/article/details/126222999)

2. 将模型下载到本地使用时出现报错：`ModuleNotFoundError: No module named 'transformers_modules.iioSnail/ReaLiSe-for-csc'`

解决方案：将 `iioSnail/ChineseBERT-for-csc` 改为 `iioSnail\ChineseBERT-for-csc`，或升级transformers