Introduction
Source code for the paper:
Improving Grammatical Error Correction via Pre-Training a Copy-Augmented Architecture with Unlabeled Data
Authors: Wei Zhao, Liang Wang, Kewei Shen, Ruoyu Jia, Jingming Liu
Arxiv url: https://arxiv.org/abs/1903.00138
Comments: Accepted by NAACL 2019 (oral)
Dependecies
- PyTorch version >= 1.0.0
- Python version >= 3.6
Downloads
- Download CoNLL-2014 evaluation scripts
cd gec_scripts/
sh download.sh
-
Download pre-processed data & pre-trained models
pre-trained model: (Google Drive/Baidu Pan)
- url1: https://drive.google.com/file/d/1zewifHUUwvqc2F-MfDRsZFio6PlSzx2c/view?usp=sharing
- url2: https://pan.baidu.com/s/1hCwQeNFjng_0_NiViJq6fg (code: mxrf)
pre-processed data: (Google Drive)(train/valid/test)
Train with the pre-trained model
cd fairseq-gec
pip install --editable
sh train.sh \${device_id} \${experiment_name}
Train without the pre-trained model
Modify train.sh to train without the pre-trained model
- delete parameter "--pretrained-model"
- change the value of "--max-epoch" to 15 (more epochs are needed without pre-trained parameters)
Evaluate on the CoNLL-2014 test dataset
sh g.sh \${device_id} \${experiment_name}
Get pre-trained models from scratch
We have public our pre-trained models as mentioned in the downloads part. We list the steps here, in case someone want to get the pre-trained models from scratch.
1. # prepare target sentences using one billion benchmark dataset
2. sh noise.sh # generate the noised source sentences
3. sh preprocess_noise_data.sh # preprocess data
4. sh pretrain.sh 0,1 _pretrain # pretrain
Acknowledgments
Our code was modified from fairseq codebase. We use the same license as fairseq(-py).
Citation
Please cite as:
@article{zhao2019improving,
title={Improving Grammatical Error Correction via Pre-Training a Copy-Augmented Architecture with Unlabeled Data},
author={Zhao, Wei and Wang, Liang and Shen, Kewei and Jia, Ruoyu and Liu, Jingming},
journal={arXiv preprint arXiv:1903.00138},
year={2019}
}