Skip to content

Latest commit

 

History

History
executable file
·
91 lines (58 loc) · 4.02 KB

README.md

File metadata and controls

executable file
·
91 lines (58 loc) · 4.02 KB
Raw

Cross-sentence Grammatical Error Correction

This repository contains the code and models to train and test cross-sentence grammatical error correction models using convolutional sequence-to-sequence models.

If you use this code, please cite this paper:

@InProceedings{chollampatt2019crosent,
  author    = {Shamil Chollampatt and Weiqi Wang and Hwee Tou Ng},
  title     = {Cross-Sentence Grammatical Error Correction},
  booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
  year      = {2019}
}

Prerequisites

Data Processing:

  • Python 2.7
  • To generate exactly the excat same data with same tokenization, you may require NLTK v2.0b7 and LangID.py v1.1.6.

Training Baseline and CroSent models:

  • Python 3.6
  • PyTorch 0.4.1

Training, and running rescorer:

  • Python 2.7
  • Moses v3

For training NUS3 models:

  • Fast_align, Moses: for computing word alignments for edit weighted log likelihood loss

Decoding using pre-trained cross-sentence GEC models

  1. Run prepare_test.sh to prepare the test datasets.

  2. Download all pre-requiste components (BPE model, embeddings, dictionaries, and pre-trained decoder) using the download.sh

  3. Download CroSent models using download_pretrained_crosent.sh script.

  4. Decode development/test sets with decode.sh.

./decode.sh $testset $modelpath $dictdir $optionalgpu

$testset is the test dataset name. The test dataset files are in the format data/test/$testset/$testset.tok.src (for the input source sentences) and data/test/$testset/$testset.tok.ctx (for the context sentences, i.e. 2 previous sentences per line)

$modelpath: could be a file for decoding using a single model or a directory for ensemble (any model with the name checkpoint_best.pt within the specified directory will be used in the ensemble). If single model, the decoder will output the files into a directory in the same location as the model path, with the name same as the model path with a prefix outputs.. If ensemble, the decoder will output the files into outptus/ directory within $model_path

$dictdir contains the path to the dictionaries. For pre-trained models it is models/dicts

$optionalgpu is an optional parameter indicating GPU id to run the decoding on (default=0).

  1. Run rearnker using the downloaded weights:
./reranker_run.sh $outputsdir $testset $weightsfile $optionalgpu

where $outputsdir is the directory which contains the output of the decoding and $weightsfile is the paths to trained weights (in the case of pretrained weights, it is models/reranker_weights/weights.nucle_dev.txt)

Training from scratch

Data preparation

Download the required datasets and run prepare_data.sh with the paths to Lang-8 and NUCLE to prepare the datasets.

Training

Download all pre-requiste components (BPE model, dictionary files, embeddings, and pre-trained decoder) using the download.sh

Each training script train_*.sh has a parameter to specify the random seed value. To train 4 different models, run the training script 4 times by variying the seed values (e.g., 1, 2, 3, 4)

For training the baseline models use train_baseline.sh script.

For training the crosent models, use train_crosent.sh script.

For training the NUS2 model, use train_nus2.sh script.

For training the NUS3 model

  1. Generate alignments using fastalign (Requires fast_align and moses under tools/ directory), run create_alignment.py data/processed
  2. Run train_nus3.sh script.

For training the reranker:

  1. Decode development dataset using ./decode.sh (steps mentioned earlier). Set $outputsdir to the output directory of this decoding step.

  2. Run ./reranker_train.sh $outputsdir $devset $optionalgpu

License

The source code is licensed under GNU GPL 3.0 (see LICENSE) for non-commerical use. For commercial use of this code, separate commercial licensing is also available. Please contact Prof. Hwee Tou Ng ([email protected])