soft-masked-bert

An unofficial implementation of soft-masked bert by pytorch and transformers, welcome to issue and PR.

Paper: Spelling Error Correction with Soft-Masked BERT

Train data: Baidu Cloud (password:l545)

Pretrained bert model: Chinese-BERT-wwm

You can add custom data following the data format:

{'text': '但是我不能去参加，因为我有一点事情阿！', 'mistakes': [{'wrong': '阿', 'correct': '啊', 'loc': '18'}]}

You can train or inference the model by train.py

python train.py