Skip to main content

Python API for multilingual legal document classification with EuroVoc descriptors using BERT models.

Project description

IMPORTANT NOTE!

This repository is not compatible yet with the latest version of the transformers library. You need to install a version < 4.11.0. We recommend and tested with transformers==4.10.3.

We are currently working to migrate the models to the latest version of transformers.

EuroVoc-BERT

PyEuroVoc is a tool for legal document classification with EuroVoc descriptors. It supports 22 languages: Bulgarian (bg), Czech (cs), Danish (da), German (de), Greek (el), English (en), Spanish (es), Estonian (et), Finnish (fi), French (fr), Hungarian (hu), Italian (it), Lithuanian (lt), Latvian (lv), Maltese (mt), Dutch (nl), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Sweedish (sv).

The tool uses BERT at its core. The list of BERT variant for each language can be found here. The performance of each model is outlined in our paper.

Installation

Make sure you have Python3 installed and a suitable package for PyTorch. Then, install pyeurovoc with pip:

pip install pyeurovoc

Usage

Import the EuroVocBERT class from pyeurovoc. Instantiate the class with the desired langauge (default is "en") and then simply pass a document text to the model.

from pyeurovoc import EuroVocBERT

model = EuroVocBERT(lang="en")
prediction = model("Commission Decision on a modification of the system of aid applied in Italy in respect of shipbuilding")

The prediction of the model is a dictionary that contains the predicted ID descriptors as keys together with their confidence score as values.

{'1519': 0.9990228414535522, '889': 0.9199628829956055, '155': 0.8993383646011353, '5541': 0.6949614882469177, '365': 0.03358528017997742, '431': 0.03317515179514885}

The number of most probable labels returned by the model is controlled by the num_labels parameter (default is 6).

prediction = model("Commission Decision on a modification of the system of aid applied in Italy in respect of shipbuilding", num_labels=4)

Which outputs:

{'1519': 0.9990228414535522, '889': 0.9199628829956055, '155': 0.8993383646011353, '5541': 0.6949614882469177}

Training your own models

Download Dataset

Firstly, you need to download the datasets. Use the download_datasets.sh script in data to do that.

./download_datasets.sh

Preprocess

Once the datasets has finished downloading, you need to preprocess them using the preprocess.py script. It takes as input the model per language configuration file and the path to the dataset.

python preprocess.py --config [model_config] --data_path [dataset_path]

Train

Training is done using the train.py script. It will automatically load the preprocessed files created by the previous step, and will save the best model for each split at the path given by the -save_path argument. To view the full list of available arguments, run python train.py --help.

python train.py --config [model_config] --data_path [dataset_path] 
                --epochs [n_epochs] --batch_size [batch_size] 
                --max_grad_norm [max_grad_norm]
                --device [device]
                --save_path [model_save_path]
                --logging_step [logging_step]
                --verbose [verbose]

Evaluate

To evaluate the performance of each model on a split, run the evaluate.py script. As in the case of training, it provides several arguments that can be visualized with python evaluate.py --help.

python evaluate.py --config [model_config] --mt_labels [mt_labels_path] --data_path [dataset_path]
                   --models_path [models_ckpt_path] 
                   --batch_size [batch_size]
                   --device [device]
                   --output_path [results_output_path]
                   --loggin_step [logging_step]
                   --verbose [verbose]

Acknowledgments

This research was supported by the EC grant no. INEA/CEF/ICT/A2017/1565710 for the Action no. 2017-EU-IA-0136 entitled “Multilingual Resources for CEF.AT in the legal domain” (MARCELL).

Credits

Please consider citing the following paper as a thank you to the authors of the PyEuroVoc:

Avram, Andrei-Marius, Vasile Pais, and Dan Tufis. "PyEuroVoc: A Tool for Multilingual Legal Document Classification with EuroVoc Descriptors." arXiv preprint arXiv:2108.01139 (2021).

or in .bibtex format:

@article{avram2021pyeurovoc,
  title={PyEuroVoc: A Tool for Multilingual Legal Document Classification with EuroVoc Descriptors},
  author={Avram, Andrei-Marius and Pais, Vasile and Tufis, Dan},
  journal={arXiv preprint arXiv:2108.01139},
  year={2021}
}

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pyeurovoc-1.3.0.tar.gz (7.3 kB view hashes)

Uploaded Source

Built Distribution

pyeurovoc-1.3.0-py3-none-any.whl (7.6 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page