GECToR-based candidate generator

This repository contains the extension of the GECToR model for Grammatical Error Correction. It yields multiple candidate corrections instead of a single hypothesis. This allows to use this model as candidate generator on the first stage of Generate-then-Rerank approach to GEC. In particular, it is used in EditScorer, described in

Improved grammatical error correction by ranking elementary edits
Alexey Sorokin

The original GECToR model was described in

GECToR – Grammatical Error Correction: Tag, Not Rewrite
Kostiantyn Omelianchuk, Vitaliy Atrasevych, Artem Chernodub, Oleksandr Skurzhanskyi
Grammarly
15th Workshop on Innovative Use of NLP for Building Educational Applications (co-located with ACL 2020)

Installation and usage

To install the package, simply run:

pip install -r requirements.txt

The project was tested using Python 3.7.

We refer the user to the original GECToR repository for the complete documentation. Here we describe only our contribution, the candidate generator.

Candidate generator

Our extension outputs all candidate edits whose probability is not less than P(KEEP) + Θ, where KEEP is the "do nothing" operation and Θ is the predefined threshold (-3.0 in the examples below).

To produce the output for the BEA2019 train, dev and test file, run the following commands:

python custom/generate_variants.py -t -3.0 -o TRAIN_OUTPUT_PATH -i DATA_DIR/wi+locness/m2/ABC.train.gold.bea19.m2 -S
python custom/generate_variants.py -t -3.0 -o DEV_OUTPUT_PATH -i DATA_DIR/wi+locness/m2/ABCN.dev.gold.bea19.m2 -l 1 -L 1000 -S
python custom/generate_variants.py -t -3.0 -o TEST_OUTPUT_PATH -i DATA_DIR/wi+locness/test/ABCN.test.bea19.orig -l 1 -L 1000 -S -r

Here DATA_DIR is where you unzip the archive with BEA2019 data. Command line options mean the following:

• -t: the threshold Θ defined above;
• -S: whether to split multiword edits to shorter elementary parts;
• -r: whether the input is a raw file;
• -l: minimal length of a sentence to be processed;
• -L: maximal length of a sentence to be processed.

Citation

If you find this work is useful for your research, please cite our paper:

@inproceedings{sorokin-2022-improving,
    title = "Improved Grammatical Error Correction by Ranking Elementary Edits",
    author = "Alexey Sorokin",
    year = "2022",
    comment = "To appear in EMNLP 2022 Proceedings"
}

Also cite the original GECToR work:

@inproceedings{omelianchuk-etal-2020-gector,
    title = "{GECT}o{R} {--} Grammatical Error Correction: Tag, Not Rewrite",
    author = "Omelianchuk, Kostiantyn  and
      Atrasevych, Vitaliy  and
      Chernodub, Artem  and
      Skurzhanskyi, Oleksandr",
    booktitle = "Proceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications",
    month = jul,
    year = "2020",
    address = "Seattle, WA, USA → Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.bea-1.16",
    pages = "163--170",
    abstract = "In this paper, we present a simple and efficient GEC sequence tagger using a Transformer encoder. Our system is pre-trained on synthetic data and then fine-tuned in two stages: first on errorful corpora, and second on a combination of errorful and error-free parallel corpora. We design custom token-level transformations to map input tokens to target corrections. Our best single-model/ensemble GEC tagger achieves an F{\_}0.5 of 65.3/66.5 on CONLL-2014 (test) and F{\_}0.5 of 72.4/73.6 on BEA-2019 (test). Its inference speed is up to 10 times as fast as a Transformer-based seq2seq GEC system.",
}