README.md

---
tags:
- text2text-generation
- Chinese
- seq2seq
- grammar
language: zh
license: apache-2.0
---
# Pseudo-Native-BART-CGEC

This model is a cutting-edge CGEC model based on [Chinese BART-large](https://huggingface.co/fnlp/bart-large-chinese).
It is trained with HSK and Lang8 learner CGEC data (about 1.3M) and human-annotated training data for the exam domain.
More details can be found in our [Github](https://github.com/HillZhang1999/NaSGEC) and the [paper](https://arxiv.org/pdf/2305.16023.pdf).

## Usage

pip install transformers

```
from transformers import BertTokenizer, BartForConditionalGeneration, Text2TextGenerationPipeline
tokenizer = BertTokenizer.from_pretrained("HillZhang/real_learner_bart_CGEC_exam")
model = BartForConditionalGeneration.from_pretrained("HillZhang/real_learner_bart_CGEC_exam")
encoded_input = tokenizer(["北京是中国的都。", "他说：”我最爱的运动是打蓝球“", "我每天大约喝5次水左右。", "今天，我非常开开心。"], return_tensors="pt", padding=True, truncation=True)
if "token_type_ids" in encoded_input:
    del encoded_input["token_type_ids"]
output = model.generate(**encoded_input)
print(tokenizer.batch_decode(output, skip_special_tokens=True))
```

## Citation

```
@inproceedings{zhang-etal-2023-nasgec,
    title = "{Na}{SGEC}: a Multi-Domain Chinese Grammatical Error Correction Dataset from Native Speaker Texts",
    author = "Zhang, Yue  and
      Zhang, Bo  and
      Jiang, Haochen  and
      Li, Zhenghua  and
      Li, Chen  and
      Huang, Fei  and
      Zhang, Min"
    booktitle = "Findings of ACL",
    year = "2023"
    }
```