This repository contains the following directories:
This directory contains the Python notebook (evaluation.ipynb) and file (evaluate_functions.py) used to evaluate the predictions for each trained model.
The different subdirectories contain three .txt files; these are the predictions for that specific model on the three test novels.
This directory contains the separate gold novels used for the validation and test set.
The Annotation_Guidelines_19th_Dutch_Spelling.pdf file contains the guidelines for annotating or correcting 19th-century Dutch novels.
This directory contains the automatically annotated novels as .txt files for the 5k and 10k finetuning datasets.
The training_data10k.txt and training_data5k.txt are the novels combined and cleaned.
The literature_list.txt contains the names of all the novels taken from Project Gutenberg.
This directory contains the code used for finetuning and testing the different T5 versions.
Where T5Trainer and T5Predict are used to train and test the models.
Example:
To run the T5Trainer script for finetuning ByT5 with a custom set of parameters, use the command:
python3 T5Trainer.py -tf google/byt5-small -train training_data.txt -dev validation_data.txt -lr 5e-5 -bs 32 -sl_train 155 -sl_dev 455 -ep 20 -es val_accuracy -es_p 3
See T5Trainer for a full explanation of each of the arguments.
To run the T5Predict with a trained ByT5 model on the three gold novels, use the following command:
python3 T5Predict.py -tf 'google/byt5-small' -weights 'byt5-small_weights.h5' -temp 0.4 -cs 40 -n_beam 2 -test1 'Multatuli_MaxHavelaar.txt' -test2 'ConanDoyle_SherlockHolmesDeAgraSchat.txt' -test3 'Nescio_Titaantjes.txt'
See T5Predict for a full explanation of each of the arguments.
Gutenberg_Data.ipynb is the notebook used to extract and download novels from Project Gutenberg.
This notebook uses the Dutch Literature Pipeline repository from van Cranenburgh.
This directory contains the code used to further pretrain ByT5.
The pretrain_jax.py is the Python file to pretrain the model, which uses custom (hyper)parameters; see run_t5_mlm_flax.py for further explanation.
Example:
To run the pretrain_jax.py script for pretraining ByT5, use the following command:
python3 pretrain_jax.py \
--output_dir "." \
--model_name_or_path "google/byt5-small" \
--model_type "t5" \
--config_name "google/byt5-small" \
--tokenizer_name "google/byt5-small" \
--do_train --do_eval \
--adafactor \
--train_file "pretraining_train.txt" \
--validation_file "pretraining_dev.txt" \
--max_seq_length "512" \
--per_device_train_batch_size "8" \
--per_device_eval_batch_size "8" \
--learning_rate "0.0005" \
--overwrite_output_dir \
--num_train_epochs "25" \
--logging_steps "10000" \
--save_steps "50000" \
--eval_steps "10000" \
--weight_decay "0.001" \
--warmup_steps "10000" \
--mean_noise_span_length "20" \
--grad_acc "1"
the train_file and validation_file arguments are where the training and validation files are given as input, which can be created with the create_data.ipynb notebook.
The create_data.ipynb notebook is used to convert Dutch data into a single train and validation file.
The convert_jax_to_torch.py is the script used to convert the pretrained weights in Flax format into a Pytorch format.
A pretrained and finetuned model is available at https://huggingface.co/AWolters/ByT5_DutchSpellingNormalization
Scripts to run the evaluation on downstream tasks. See the README for more information.
If you use this data or model for research, please cite the paper:
@article{wolters2022neuralspellnorm,
author={Wolters, Andre and van Cranenburgh, Andreas},
year={2023},
title={Historical Dutch Spelling Normalization with Pre-trained Language Models},
journal={Computational Linguistics in the Netherlands Journal},
volume={13},
pages={147--171},
url={https://clinjournal.org/clinj/article/view/178},
}