https://github.com/ACL2020SpellGCN/SpellGCN

SpellGCN
https://github.com/ACL2020SpellGCN/SpellGCN

Last synced: 3 months ago
JSON representation

SpellGCN

• Host: GitHub
• URL: https://github.com/ACL2020SpellGCN/SpellGCN
• Owner: ACL2020SpellGCN
• Created: 2019-12-09T15:02:54.000Z (almost 5 years ago)
• Default Branch: master
• Last Pushed: 2021-02-28T09:22:00.000Z (over 3 years ago)
• Last Synced: 2024-06-24T05:50:01.187Z (5 months ago)
• Language: Python
• Homepage: https://arxiv.org/abs/2004.14166
• Size: 25.1 MB
• Stars: 240
• Watchers: 8
• Forks: 63
• Open Issues: 17
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• StarryDivineSky - ACL2020SpellGCN/SpellGCN

README

        
# SpellGCN

[SpellGCN](https://arxiv.org/abs/2004.14166) is a method for chinese spelling check, which embeds the visual and phological knowledge into BERT.

This repository contains data, evaluation and training scripts.

Citation:

```

@inproceedings{DBLP:journals/corr/abs-2004-14166,

  author    = {Xingyi Cheng and

               Weidi Xu and

               Kunlong Chen and

               Shaohua Jiang and

               Feng Wang and

               Taifeng Wang and

               Wei Chu and

               Yuan Qi},

  title={SpellGCN: Incorporating Phonological and Visual Similarities into

               Language Models for Chinese Spelling Check},

  booktitle={ACL},

  year={2020}

}

```

This is the official code for paper titled "SpellGCN: Incorporating Phonological and Visual Similarities into Chinese Spelling Check".

## How to run?

The code is based on Tensorflow==1.13.1 and python 2.7 or higher

Run commands as follows:

```

cd scripts/

conda create -n spellgcn python=2.7.1

source activate spellgcn

pip install tensorflow==1.13.1

sh run.sh

```

Note: Since SpellGCN is based on BERT, the path to the BERT directory should be provided in the run.sh.

The default training data is the combination of data samples from SIGHAN13, SIGHAN14, SIGHAN15.

The additional 270K data samples are absent here due to the lack of permission.

## Contact

[email protected] and [email protected]