kanekomasahiro / bert-gec

Encoder-Decoder Models Can Benefit from Pre-trained Masked Language Models in Grammatical Error Correction

Code for the paper: "Can Encoder-Decoder Models Benefit from Pre-trained Language Representation in Grammatical Error Correction?" (In ACL 2020). If you use any part of this work, make sure you include the following citation:

@inproceedings{Kaneko:ACL:2020,
    title={Encoder-Decoder Models Can Benefit from Pre-trained Masked Language Models in Grammatical Error Correction},
    author={Masahiro Kaneko, Masato Mita, Shun Kiyono, Jun Suzuki and Kentaro Inui},
    booktitle={Proc. of the 58th Annual Meeting of the Association for Computational Linguistics (ACL)},
    year={2020}
}

Requirements

• python >= 3.5
• torch == 1.1.0
• bert-nmt
• subword
• gec-pseudodata

How to use

• First download the necessary tools using the following command:

cd scripts
./setup.sh

• This code uses wi+locness dataset.
• Note that since the gold of wi+locnness test data is not available, validation data was specified as test data.
• Place your data in the data directory if necessary.
• You can train the BERT-GEC model with the following command:

./train.sh

• You can use trained BERT-GEC model with the following command:
• This model achieves the F score 62.77 on CoNLL.
• The results in the paper are initialized with four pre-trained models with different seeds.

./generate.sh /path/your/data gpu

• The OUTPUT file contains the system outputs of ensembled models.

License

See the LICENSE file