Interpretability for Language Learners Using Example-Based Grammatical Error Correction

Code for the paper: "Interpretability for Language Learners Using Example-Based Grammatical Error Correction" (In ACL 2022). If you use any part of this work, make sure you include the following citation:

@inproceedings{Kaneko:ACL:2022,
    title={Interpretability for Language Learners Using Example-Based Grammatical Error Correction},
    author={Masahiro Kaneko, Sho Takase, Ayana Niwa, Naoaki Okazaki},
    booktitle={Proc. of the 60th Annual Meeting of the Association for Computational Linguistics (ACL)},
    year={2022}
}

Setup

We use Python version == 3.7.10.

All requirements can be found in requirements.txt. You can install all required packages with following:

pip install -r requirements.txt

You can also install fairseq with following:

cd knnmt
pip install --editable ./

You need to place train.src, train.trg, dev.src, dev.trg, test.src in data directory.

To train EB-GEC model

You can train EB-GEC model using train.sh in scripts directory.

cd scripts
./train.sh $seed

To generate corrected texts and examples with EB-GEC model

You can generate corrected texts and examples with EB-GEC model using generate.sh in scripts directory.

cd scripts
./generate.sh $seed

In the output nbest file, SRC_EXAMPLE and TGT_EXAMPLE for EDIT (. → :) for source S and detokenized output D are displayed as follows:

S-527 There are three reasons as follows .
D-527 -0.24829396605491638  There are three reasons as follows :
EDIT-527  :
SRC_EXEMPLE-527 there are two big reasons as follows .
TGT_EXAMPLE-527 There are two big reasons as follows :