Nakdimon: a simple Hebrew diacritizer

Repository for the paper Restoring Hebrew Diacritics Without a Dictionary by Elazar Gershuni and Yuval Pinter.

Demo: https://nakdimon.org/

Requires Python 3.12+. The runtime uses ONNX Runtime (no TensorFlow required to run inference).

Locally:

$ pip install nakdimon
$ diacritize input_file.txt -o=output_file.txt

Building and running docker container

$ docker build -t nakdimon .
$ docker run --rm -it nakdimon /bin/bash

Development setup (with uv)

$ uv sync                  # install runtime deps
$ uv sync --extra train    # add TensorFlow stack (Python 3.12–3.13 only)
$ uv sync --extra research # add matplotlib/seaborn for plots

Training and evaluating

Training requires the [train] extra (TensorFlow + wandb + tf2onnx):

$ pip install 'nakdimon[train]'

Then:

> python -m nakdimon train --model=models/Nakdimon.keras
> python scripts/convert_to_onnx.py models/Nakdimon.keras models/Nakdimon.onnx
> python -m nakdimon run_test --test_set=tests/new --model=models/Nakdimon.onnx
> python -m nakdimon results --test_set=tests/new --systems Snopi Morfix Dicta MajAllWithDicta Nakdimon

The trained .h5 is converted to .onnx once; the runtime predictor consumes .onnx. By default, the bundled model is nakdimon/data/Nakdimon.onnx (shipped in the wheel).

The second step asks the Nakdimon server to predict the diacritics for the test set. You may skip this step. A folder for the results is created in the chosen test folder, with the same name as the model; in this case, tests/new/NakdimonNew. By default, the test set is the one used in the paper (tests/new); you can use tests/dicta instead. If the test results already exist, you may skip this step. If you are not sure, you can use the --skip_existing flag.

The third step calculates and prints the results (DEC, CHA, WOR and VOC metrics, as well as OOV_WOR and OOV_VOC). By default, the systems are the folders in the chosen test folder. For the Dicta test set (/tests/dicta) you should use MajAllNoDicta instead of MajAllWithDicta, otherwise the vocabulary for the Majority would include the test set itself.

Diacritizing a single file

> python nakdimon predict input_file.txt output_file.txt

Using other systems

You can use the run_test command to run the test set on other systems, such as Dicta:

> python nakdimon run_test --test_set=tests/new --system=Dicta

This will create a folder named Dicta for the results in the tests/new folder. Note that Morfix cannot be used in this manner, as its license prohibit automatic use.

Running ablation tests

You can use the --ablation flag to train different models for the ablation tests and other experiments:

> python -m nakdimon train --model=models/SingleLayer.keras --ablation=SingleLayer

See the file ablation.py for the list of available ablation parameters.

Important folders

• hebrew_diacritized is the training set.
• tests contains three tests sets: new, dicta and validation. Each test set has an expected folder that describes the ground truth. The results of python nakdimon run_test are stored in sibling folder, named after the model.
• models contains the trained model.
• nakdimon holds the source code.

Citation

@inproceedings{gershuni2022restoring,
  title={Restoring Hebrew Diacritics Without a Dictionary},
  author={Gershuni, Elazar and Pinter, Yuval},
  booktitle={Findings of the Association for Computational Linguistics: NAACL 2022},
  pages={1010--1018},
  year={2022}
}

Gershuni, Elazar, and Yuval Pinter. "Restoring Hebrew Diacritics Without a Dictionary." Findings of the Association for Computational Linguistics: NAACL 2022. 2022.