script conversion systems



Rababa is an automated diacriticization library for Abjad scripts which was originally developed for pointing Arabic text.

This is a technical blog about the excellent results we obtained when applying the “Rababa strategy and architecture” to Hebrew.

We explain how we have managed to half the decision errors recorded on the automated diacritization of Abjad scripts.

Role of diacritization

Modern Hebrew is typically written using the Hebrew alphabet that only specifies consonants, leaving vowels to be left to the reader to infer.

The process of “diacritization”, “pointing” or “Niqqud” is the Hebraic system used to specify these vowels and is found in liturgic Hebrew and taught in schools.

The “cantillation” is another set of symbols added to Hebrew texts within the religious context used to indicate how Hebrew should be chanted. Cantillation is not covered here.

Example of Hebraic diacritization (blue) with cantillation (red)
Figure 1. Example of Hebraic diacritization (blue) with cantillation (red)

Applications in informatics

Pointed Hebrew is useful and essential for applications including text-to-speech and transliteration.


The Niqqud is the set of symbols used to enrich Hebrew and specify the vocalization.

We present here the classification used we have adopted from Nakdimon (


sh’va, reduced segol, reduced patakh, reduced kamatz, hirik, tzeire, segol, patakh, kamatz, holam, kubutz, shuruk, meteg


rafe, dagesh letter, mappiq


rafe, shin yemanit, shin smalit

The rationale of the above decomposition is presented as follows:

  • diacritics on Sin are present only on the Sin letter of the alphabet;

  • the Dagesh diacritics are placed in the middle of the letters; and

  • the members of Niqqud present themselves as “dots” on the top or bottom of the alphabets.

As a result, only one member of each of the above group can enrich a letter. But a letter could, for instance, be decorated with one member of Niqqud and Dagesh classes each. Therefore, the above decomposition allows to represent all the symbols in three independent spaces.

Diacritization in Interscript


Interscript provides mappings that allow the transliteration of languages into various writing systems.

In this context, Abjad scripts like Hebrew need to be processed via several steps:

  1. Hebraic text

  2. diacritization of Hebraic text

  3. transliteration of diacriticized Hebraic text

Diacritization with deep learning

Correct diacritization requires an accurate understanding not only of the language morphemes and their variants but also the language grammar.

Furthermore, given the possible multiple meanings available to a particular word in Hebrew (called collisions), some understanding of the context is required.


This hard problem has been approached in various ways with an evolution quite typical:

  1. Rule-based approaches

  2. Machine learning approaches

  3. Deep learning approaches

For more details, we have reviewed the latest publications, tested the latest codebases and summarized the latest research ideas here.

Literature review


Nakdan is a live system based on 3 steps combining engineered linguistic information with a trained neural model:

  1. Parts-of-Speech Tagging

  2. Filtering the possible diacritizations

  3. Ranking the possible diacritizations for each word, in context


Nakdimon is a lighter system attempting to achieve diacritization by using a more powerful Neural Network model only.

The authors have also published a new data set, codes for NLP and metrics that we have largely adopted in our work.

Architecture of Rababa

Our architecture is based on the Tacotron and CBHGs, as explained in our recent blog on Arabic diacritization.

Training and results

NLP of Hebrew diacritics

We have provided a heuristic explanation of the decomposition of Diacritics as: Niqqud, Sin, and Dagesh.

We also found much better results compared to a naïve attempt where all the diacritics would co-exist within the same space.

Modelling of Niqqud, Sin and Dagesh

Compared to the architecture described in our previous blog, the simplest change was to just add 2 additional CBHG projections to the model.

The model is then trained with back-propagation in a serial fashion from the Niqqud/Dagesh and Sin projection losses (see Hebrew code).


The original dataset was adopted from the Hebrew Diacritized repository.

The dataset contains a range of diacritized texts of multiple origins, including but not limited to ancient text, religious text, modern text and poetry.

The datasets needed further cleaning, and we will publish those cleaned datasets soon.

Training strategies

Code to run experiments

We have integrated the code with Wandb to make it simpler to run extensive experiments and monitor/show results in real time.

Experiments with datasets

The variety and diacritization quality within the datasets allowed to run multiple experiments.

We found that to pre-training Rababa first with various datasets before using the modern Hebrew corpus as target would slightly improve the results.

This will discussed in more details very soon.

Hyperparameter tuning

On the top of the datasets, various parameters can be fine-tuned.

We have tried and evaluated various combinations, which will also be discussed in more details very soon.

System evaluation and performance

The following acronyms are used to describe performance of the Hebrew diacriticization system:

  • DEC: decision accuracy (%)

  • CHA: character accuracy (%)

  • WOR: word accuracy (%)

  • VOC: vocalization accuracy (%)

In order to make sure our metrics are correct, we have tested the Nakdimon code and confirmed that we can reproduce identical results using their test dataset.

Scores after training




















Values for Nakdimon are reproduced from Table 3 of the Nakdimon paper.

In these results we compare the implementations of Nakdan which provided the best published results from a hybrid system (NNets + rules + search), Nakdimon which only uses NNets, against Rababa.

The evidence of these values can be found in exploring the following link.


Not only we could adapt Rababa to the problem of Hebrew diacritization, but using the good work made on NLP, modelling and the datasets published by other team, we could surpass the benchmarks by aa substantial margin.

We confirm that as mentioned in the 2021 Nakdimon paper, that the results of Rababa could be improved by pre-training on ancient or religious datasets prior to targeting a smaller, modern one.

While Rababa’s decision accuracy error (100% - DER) of 0.37% is to be compared with the previous best accuracy error of 1.04% for hybrid systems in Hebrew, the corresponding performance using Rababa for Arabic diacritization using a similar metric (100% - DER*) is currently around 0.85%.


Rababa produces a 65% improvement over the previous best automated diacritization of Hebrew, held by a hybrid system, thus entering into a new domain for what deep learning can achieve with Abjad diacriticization.