ECSpell

Code for paper "General and Domain Adaptive Chinese Spelling Check with Error Consistent Pretraining"

📖 Introduction

This repository contains the official PyTorch implementation for ECSpell. Our approach leverages error-consistent pretraining to improve generalization and domain adaptation for Chinese Spelling Check (CSC).

🛠️ Environment Setup

We recommend Python 3.8+ and PyTorch >= 1.9.0.

conda create -n ecspell python=3.8 -y
conda activate ecspell
pip install torch transformers jieba pypinyin Pinyin2Hanzi tqdm scipy

Install the local glyce library:

cd glyce
python setup.py develop
pip show glyce   # to ensure the successful installation of glyce lib

📁 Data Preparation

1. Model Weights: Download the pre-trained weights from Google Drive and place them in Transformers/glyce/.
2. Datasets: Download SIGHAN or other CSC datasets. Organize them following the structure:
   • Data/domains_data/ for domain-specific data.
   • data/basedata/simplified/ or data/basedata/traditional/ for baseline training data.
3. Tokenization (if required):

   python3 Code/tokenize_sequence.py
   

🏋️ Training

Set the Python path and run the training script:

export PYTHONPATH=$PYTHONPATH:./:./Code
python Code/train_baseline.py \
  --config_path Transformers/glyce_bert_both_font.json \
  --output_dir ./outputs

(Adjust arguments in train_baseline.py as needed for your specific setup.)

📊 Evaluation & Inference

• For evaluation, use the scripts in csc_evaluation/ or modify the training script to enable evaluation mode.
• csc_evaluation/build_train_data.py can be used to construct and split training/validation datasets.

📝 Data Usage Notes

Path: Data/domains_data

• For zero-shot tasks, you should combine the *.train file and *.test file.
• For common tasks, the *.train file is used to do training and do evaluating while *.test is adopted to do predicting.

📜 Citation

If you use this code in your research, please cite our paper:

@article{lv2023general,
  title={General and Domain-adaptive Chinese Spelling Check with Error-consistent Pretraining},
  author={Lv, Qi and Cao, Ziqiang and Geng, Lei and Ai, Chunhui and Yan, Xu and Fu, Guohong},
  journal={ACM Transactions on Asian and Low-Resource Language Information Processing},
  volume={22},
  number={5},
  pages={1--18},
  year={2023},
  publisher={ACM New York, NY}
}