/oteann2

Primary LanguageJupyter NotebookMIT LicenseMIT

OTEANN

This repository is the official implementation of OTEANN: Estimating the Transparency of Orthographies with an Artificial Neural Network.

Requirements

  • Require Ubuntu 18.04, Python 3.6+.

  • Download files on your machine

  • Go to the oteann2 main directory

    • cd oteann2

  • Create a virtual environment

    • python3 -m venv oteann

  • Activate virtual environment

    • source oteann/bin/activate

  • Load the python librairies needed for OTEANN (e.g. numpy, pandas, torch...) from the requirements file

    • python -m pip install -r requirements.txt

  • Download the subdatasets (required local free spacesize=334Mo)

    • python download_subdatasets.py

Evaluation and Training

In order to run this code, you need to:

  • Run the oteann.ipynb in order to create the ANN model, train it, test it and display the results of the paper.

Results

OTEANN achieves the following performance:

orthography write score read score
ent 99.4 ± 0.2 99.3 ± 0.3
eno 00.0 ± 0.0 00.0 ± 0.0
ar 83.8 ± 1.3 99.4 ± 0.3
br 80.0 ± 1.1 76.3 ± 1.7
de 68.9 ± 1.4 78.2 ± 1.0
en 36.5 ± 0.6 33.0 ± 1.8
es 67.0 ± 1.3 85.2 ± 1.2
fi 97.9 ± 0.4 92.8 ± 0.8
fr 27.8 ± 0.9 78.3 ± 1.2
fro 99.1 ± 0.3 89.0 ± 1.2
it 94.9 ± 0.8 72.0 ± 1.4
ko 82.0 ± 1.2 97.7 ± 0.5
nl 73.1 ± 1.8 56.6 ± 1.5
pt 70.7 ± 1.2 32.2 ± 1.8
ru 42.4 ± 1.7 97.6 ± 0.7
sh 99.1 ± 0.2 54.4 ± 2.3
tr 95.5 ± 0.7 96.0 ± 0.7

License

OTEANN uses minGTPT (https://github.com/karpathy/minGPT) which is released under the MIT licence.

Citation

This code was used for the following paper:

@misc{marjou2020oteann,
      title={OTEANN: Estimating the Transparency of Orthographies with an Artificial Neural Network}, 
      author={Xavier Marjou},
      year={2020},
      eprint={1912.13321v2},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}