This repository contains the implementation for GECTurk; the first extensive dataset and analysis on Grammatical Error Detection (GED) and Grammatical Error Correction (GEC) for Turkish.
Figure 1: Data generation pipeline.
We present a flexible and extensible synthetic data generation pipeline for Turkish covering 25 expert-curated grammar rules implemented through transformation functions.
The first step in this pipeline involves obtaining a grammatically correct sentence from an existing corpus. Following this, morphological analysis is conducted on the sentence. Next, the validity of the sentence for the transformation function is examined. If the sentence is deemed suitable, the transformation is implemented with a certain probability 'p'. The next phase involves selecting tokens to modify, after which the pipeline checks if reverse transformations can retrieve the original tokens. In case the original cannot be recovered, that particular sentence is dismissed. However, if the original can be recovered, the transformed sentence is consequently added to the corpus.
GECTurk consists of 138K sentences, consisting of 25 different rules, with varying levels of frequency. Here are some samples from GECTurk dataset:
S Meclis'i boykot kadar anlamsız bir protesto ola bilir mi?
A 6 8|||rule_20|||olabilir|||REQUIRED|||-NONE-|||0
Meclis'i boykot kadar anlamsız bir protesto olabilir mi?
S Hiç bir haber hiç bir gazeteciye gökten inmez.
A 0 2|||rule_22|||Hiçbir|||REQUIRED|||-NONE-|||0
A 3 5|||rule_22|||hiçbir|||REQUIRED|||-NONE-|||0
Hiçbir haber hiçbir gazeteciye gökten inmez.
To further evaluate, we manually annotate a dataset of movie reviews which is also provided in here and at huggingface repository.
In our work, we focus on two tasks: Grammatical Error Correction (GEC) and Grammatical Error Detection (GED).
We use Neural Machine Translation (NMT) (only for GEC), masked Generative Pre-trained Transformer (mGPT) using prefix-tuning, and sequence tagger pre-trained on Turkish to evaluate the performance of GECTurk.
Here are the necessary instructions to download and use the datasets, reproduce results, and continue generating data.
Datasets
The datasets can be downloaded from here or from huggingface repository. After downloading the dataset, move it to the corresponding dataset directory.
We also provide preprocessed version of the "de/da clitic errors" dataset proposed by Arikan et. al. in the datasets/boun directory. They follow the M2 format and annotated with the error classes proposed by the GECTurk paper.
mkdir datasets
mv <path_to_dataset> datasets/<dataset_name>Where dataset_name is either gec_tr or boun.
Weights
We provide the weights trained on all data for the NMT, mGPT and sequence-tagger in here. Our best performing model, sequence-tagger can also be accessed from here. After obtaining the weights, move it to the corresponding folder in the weights directory under the model name.
mkdir weights weights/<model_name>
mv <weight> weightswhere model_name can be nmt, mgpt or sequence_tagger.
For each model, a default configuration file is provided in the configs directory. You can change the parameters in the configuration file or create a new one.
NMT Training and Validation
To train or validate the models, after downloading the dataset and the weights (for validation) you can use the following commands:
# an example for training NMT
python nmt_based/train.py --train_config configs/nmt/train.yaml --split gec_tr# an example for testing NMT
python nmt_based/test.py --test_config configs/nmt/test.yaml --split gec_trmGPT Training and Validation
Run this from inside the mGPT folder, not the root folder. The paths for the datasets should be modified as needed inside the mgpt_generation.py file. This file will perform all the training, and generate the file for testing. To evaluate for correction, the m2_scorer tool should be used from the utils folder.
In order to use the dataset_percentage with anything other than 1.0, you will need to make your own dataset splits and name them properly.
cd mGPT
python mgpt_generation.py --dataset_percentage 1.0 --epochs 3 --weights_file weights_name.ckpt --output_file output_file.txtSequence-tagger Training and Testing
# example command for training the sequence-tagger model
python sequence_tagger/train.py --train_config configs/sequence_tagger/train.yaml# example command for testing the sequence-tagger model
python sequence_tagger/test.py --test_config configs/sequence_tagger/test.yaml
> Generating data
To produce more data, create a text file that contains grammatically correct sentences on each line. Afterward, execute the following command:
```bash
python data_generation/process_sentences.py --source path/to/your/input/sentences --experiment name/of/the/experimentIt is fairly simple. You need to implement a transformation and reverse transformation function at the data_generation/rules.py and data_generation/correction_rules.py files respectively. Each transformation function accepts 4 arguments, sentence, morphological_analysis, flags, p and name. You can benefit from the morphological analyzer to implement more sophisticated rules. Any dictionary you may need should be added to data_generation/dictionaries.py file.
You should begin by curating a list of sentences that are grammatically correct. Following this, create transformation functions pertaining to the specific grammar rules you desire, and also develop their corresponding reverse transformation functions. Lastly, incorporate a new morphological analyzer for the target language.
- Supported by the Scientific and Technological Research Council of Türkiye (TÜBİTAK) as part of the project Automatic Learning of Procedural Language from Natural Language Instructions for Intelligent Assistance with the number 121C132.
- We gratefully acknowledge KUIS AI Lab for providing computational support.
- We thank our anonymous reviewers and the members of GGLab who helped us improve this paper.
@misc{kara2023gecturk,
title={GECTurk: Grammatical Error Correction and Detection Dataset for Turkish},
author={Atakan Kara and Farrin Marouf Sofian and Andrew Bond and Gözde Gül Şahin},
year={2023},
eprint={2309.11346},
archivePrefix={arXiv},
primaryClass={cs.CL}
}For any questions or inquiries, please contact Atakan Kara.
Please send us an e-mail or report an issue, if something is broken or if you have further questions.