There are source codes for Neural Quality Estimation with Multiple Hypotheses for Grammatical Error Correction Our Paper.

• Python==3.6.2
• Pytorch==1.2.0
• transformers==3.2.0
• pip3 install errant
• Clone the jfleg project from here https://github.com/keisks/jfleg

• All these files can be downloaded and you should put them in the corresponding folders.
• All data can be found at Ali Drive.
• The checkpoints (BERT-VERNet and ELECTRA-VERNet) can be found at Ali Drive.
• All features used in reranking can be found at Ali Drive.

• VERNet inherits Hugginface Transformers, you can change codes for various pretrained language models.
• Go to the model folder and train models with BERT or ELECTRA as follow:

bash train.sh

bash train_electra.sh

• These experimental results are shown in Table 3 of our paper.
• The evaluations are the same as the evaluations of GED models.
• Go to the model folder and test BERT-VERNet model or ELECTRA-VERNet model as follow:

bash test.sh

bash test_electra.sh

• First you should go to the model folder and generate features from BERT-VERNet model or ELECTRA-VERNet model. So run the following command:

bash generate_feature.sh

bash generate_feature_electra.sh

• (Optional Stage for Learning Feature Weight) Second, you can generate ranking features with the GEC model score and VERNet score, all these results are provided in the features folder. And then we train learning-to-rank models with Coordinate Ascent and get weights of ranking features. You should go to the feature_rerank folder and run the following command:

bash train.sh

bash train_electra.sh

• Finally, if you want to test the model, you can go to the feature_rerank folder and run the following command:

bash test.sh

bash test_electra.sh

Using this command, you can rerank beam search candidates and automatically evaluate the model performance on three datasets CoNLL-2014, FCE and JFLEG. For BEA19 evaluation, you should submit the runs to their hidden test website.

The results are shown as follows.

@inproceedings{liu2021vernet,
  title={Neural Quality Estimation with Multiple Hypotheses for Grammatical Error Correction},
  author={Liu, Zhenghao and Yi, Xiaoyuan and Sun, Maosong and Yang, Liner and Chua, Tat-Seng},
  booktitle={Proceedings of NAACL},
  year={2021}
}

If you have questions, suggestions, and bug reports, please email:

[email protected]