[Mova](https://md.mova.elamur.ai/free/en/) · [this page in HTML](https://mova.elamur.ai/free/en/docs/papers/aclbulk-2024-mitelman-code-switching-back-transliteration-bilingual/)

[home](https://md.mova.elamur.ai/free/en/) / [docs](https://md.mova.elamur.ai/free/en/docs/) / [papers](https://md.mova.elamur.ai/free/en/docs/papers/) / [Aramaic](https://md.mova.elamur.ai/free/en/docs/papers/arc/) / Code-Switching and Back-Transliteration Using a Bilingual Model

# Code-Switching and Back-Transliteration Using a Bilingual Model

**Authors:** Daniel Weisberg Mitelman; [Nachum Dershowitz](https://md.mova.elamur.ai/free/en/docs/papers/authors/nachum-dershowitz/); [Kfir Bar](https://md.mova.elamur.ai/free/en/docs/papers/authors/kfir-bar/) · **Year:** 2024 · **Venue:** Findings of the Association for Computational Linguistics: EACL 2024 **Link:** [aclanthology.org/2024.findings-eacl.102](https://aclanthology.org/2024.findings-eacl.102/) · library `aclbulk-2024-mitelman-code-switching-back-transliteration-bilingual` · **Licence:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) (ACL Anthology)

[**1** citations](https://openalex.org/works?filter=cites:W7126460603) openalex · 2026-10-18[**4** citations in our library](https://md.mova.elamur.ai/free/en/docs/papers/cited-by/aclbulk-2024-mitelman-code-switching-back-transliteration-bilingual/) of the 140673 papers whose bibliographies we read[Findings of the Association for Computational Linguistics: EACL 2024](https://aclanthology.org/volumes/2024.findings-eacl/)[licence](https://creativecommons.org/licenses/by/4.0/)

- [Publisher's page](https://aclanthology.org/2024.findings-eacl.102/)
- [PDF](https://aclanthology.org/2024.findings-eacl.102.pdf)
- [DOI 10.18653/v1/2024.findings-eacl.102](https://doi.org/10.18653/v1/2024.findings-eacl.102)
- [OpenAlex](https://openalex.org/W7126460603)

**Authors (3)**

1. [Daniel Weisberg Mitelman](https://openalex.org/A5025320818) h 1 · 3 citations · 2 works
2. [Nachum Dershowitz](https://openalex.org/A5039993352) [ORCID](https://orcid.org/0000-0003-0363-2735) · h 43 · 9037 citations · 540 works
3. [Kfir Bar](https://openalex.org/A5069195550) [ORCID](https://orcid.org/0000-0002-1354-2955) · h 11 · 362 citations · 81 works

Citations and author metrics from openalex (2026-10-18); authors with the same name are not always told apart.

## What it is about

The paper handles Judeo-Arabic, Arabic written in Hebrew letters with Hebrew and Aramaic borrowings, in two steps. First, a token classifier marks the non-Arabic words. It is fine-tuned from the bilingual HeArBERT model on the Friedberg Judeo-Arabic Project corpus (about 3.9 million words, 34% of the tokens non-Arabic); it reaches 98.46% accuracy and an F1 of 97.78% for the non-Arabic label. Second, a character-level classifier maps each Hebrew letter either to an Arabic letter or to an epsilon tag, the deletion of a letter. It is trained on parallel Judeo-Arabic and Arabic texts (Kuzari, Mishnah commentary, Beliefs), aligned by a new anchor-word algorithm (align rates 95.8%, 92.2% and 74.1%), and extended with synthetic Judeo-Arabic made from Arabic philosophy texts by a letter-mapping table. On the Kuzari test set the label error rate is 1.40%, against 2.48% for the best prior system and 9.51% for a simple letter mapping.

## Ideas

- I1. Anchor-word alignment: a word that occurs once on each side, has at least the average word length, and matches within five word positions of its place in the other text is an anchor; cut both texts at anchors and match words inside each segment. Use it to align Targum Onkelos with the Hebrew text in `languages/arc-x-targum/`, the way `macula` lines up verses.
- I2. Morpheme-level language tag: an Arabic clitic on a Hebrew stem (the paper's *al-* on *maskilim*) gets its own tag, and the stem gets the other, when the stem is rare in one lexicon and frequent in the other. This is a rule for `tools/langseg`, which now cuts text only at word level, and would give mixed Hebrew and Aramaic words a split.
- I3. Epsilon as deletion: a letter table with a deletion option is a small finite-state transducer, and its letter error rate is a clean measure. Use the same form for script-mapping tables in `languages/jpa/` or `languages/tmr/` and for correction rules in `tools/ocrfix`.

## For Mova

I1 goes into `languages/arc-x-targum/` as an alignment step. I2 goes into `tools/langseg` (the varieties hbo, arc, tmr, jpa). I3 goes into `tools/ocrfix`. Judeo-Arabic is not a Mova variety; these are ideas on contact-script problems only, and the ideas are not tied to any other part of this paper.

## Cite

The paper itself, not this card. BibTeX from the library's record of the work. [DOI 10.18653/v1/2024.findings-eacl.102](https://doi.org/10.18653/v1/2024.findings-eacl.102)

```
@inproceedings{mitelman2024code,
  title = {{Code-Switching and Back-Transliteration Using a Bilingual Model}},
  author = {Mitelman, Daniel Weisberg and Dershowitz, Nachum and Bar, Kfir},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2024},
  year = {2024},
  doi = {10.18653/v1/2024.findings-eacl.102},
  url = {https://doi.org/10.18653/v1/2024.findings-eacl.102}
}
```

- [← previous A Transformer-based parser for Syriac morphology](https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2023-naaijer-transformer-based-parser-syriac-morphology/)
- [next → Modeling linking between text and lexicon with OntoLex-Lemon: a case study of computational terminology for the Babylonian Talmud](https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2024-sciolette-modeling-linking-between-text-lexicon/)
- [↑ back to Aramaic](https://md.mova.elamur.ai/free/en/docs/papers/arc/)

## References

11 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Kfir Bar, Nachum Dershowitz, Lior Wolf, Yackov Lubarsky, and Yaacov Choueka. 2015. Processing JudeoArabic texts. In Proceedings of the First Inter national Conference on Arabic Computational Lin guistics (ACLing ’15, Cairo, Egypt), pages 138–144. [find](https://scholar.google.com/scholar?q=Kfir+Bar%2C+Nachum+Dershowitz%2C+Lior+Wolf%2C+Yackov+Lubarsky%2C+and+Yaacov+Choueka%2E+2015%2E+Processing+JudeoArabic+texts%2E+In+Proceedings+of+the+First+Inter+national+Conference+on+Arabic+Computational+Lin+guist)
2. Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine Learning, pages 369–376. ACM. [↗](https://doi.org/10.1145/1143844.1143891)
3. Judah HaLevi. 2012. The Kuzari – The Book of Refu tation and Proof on the Despised Faith. AlKamel Verlag, Freiberg. Transliterated and edited by Nabih Bashir with assistance of ‘Abed ’lSalam Muosa. [find](https://scholar.google.com/scholar?q=Judah+HaLevi%2E+2012%2E+The+Kuzari+%E2%80%93+The+Book+of+Refu+tation+and+Proof+on+the+Despised+Faith%2E+AlKamel+Verlag%2C+Freiberg%2E+Transliterated+and+edited+by+Nabih+Bashir+with+assistance+of+%E2%80%98Abed+%E2%80%99lSalam+Muosa%2E)
4. Benjamin Hary. 2018. JudeoArabic in the Arabic speaking world. In B. Hary and S. Benor, editors, Languages in Jewish Communities, Past and Present, pages 35–69. de Gruyter, Boston. [find](https://scholar.google.com/scholar?q=Benjamin+Hary%2E+2018%2E+JudeoArabic+in+the+Arabic+speaking+world%2E+In+B%2E+Hary+and+S%2E+Benor%2C+editors%2C+Languages+in+Jewish+Communities%2C+Past+and+Present%2C+pages+35%E2%80%9369%2E+de+Gruyter%2C+Boston%2E)
5. Adina Hoffman and Peter Cole. 2011. Sacred Trash: The Lost and Found World of the Cairo Geniza. Schocken, New York. [find](https://scholar.google.com/scholar?q=Adina+Hoffman+and+Peter+Cole%2E+2011%2E+Sacred+Trash%3A+The+Lost+and+Found+World+of+the+Cairo+Geniza%2E+Schocken%2C+New+York%2E)
6. Ashti Afasyaw Jaf and Sema Koç Kayhan. 2021. Machinebased transliterate of Ottoman to Latin based script. Scientific Programming, 2021:1–8. [find](https://scholar.google.com/scholar?q=Ashti+Afasyaw+Jaf+and+Sema+Ko%C3%A7+Kayhan%2E+2021%2E+Machinebased+transliterate+of+Ottoman+to+Latin+based+script%2E+Scientific+Programming%2C+2021%3A1%E2%80%938%2E)
7. Sarvnaz Karimi, Falk Scholer, and Andrew Turpin. 2011. Machine transliteration survey. ACM Comput. Surv., 43(3). *Table 4: Results per letter, sorted by F1 score.* behaviors. While this choice omitted some infor mative Arabic characters, future work will employ various language models that include these marks. [↗](https://doi.org/10.1145/1922649.1922654)
8. Lorenz Nigst, Maxim Romanov, Sarah Bowen Sa vant, Masoumeh Seydi, and Peter Verkinderen. 2023. OpenITI: A machinereadable corpus of Islamicate texts. Zenodo. We see no potential ethical issues in this work. [find](https://scholar.google.com/scholar?q=Lorenz+Nigst%2C+Maxim+Romanov%2C+Sarah+Bowen+Sa+vant%2C+Masoumeh+Seydi%2C+and+Peter+Verkinderen%2E+2023%2E+OpenITI%3A+A+machinereadable+corpus+of+Islamicate+texts%2E+Zenodo%2E+We+see+no+potential+ethical+issues+in+this)
9. Yoav Phillips. 2020. Identifying the Islamic sources of baḥya ibn paqūda’s: “kitāb alhidāya ʾilā farāʾiḍ alqulūb”: Using automatic transliteration and text reuse processes. Master’s thesis, Haifa University, July. [find](https://scholar.google.com/scholar?q=Yoav+Phillips%2E+2020%2E+Identifying+the+Islamic+sources+of+ba%E1%B8%A5ya+ibn+paq%C5%ABda%E2%80%99s%3A+%E2%80%9Ckit%C4%81b+alhid%C4%81ya+%CA%BEil%C4%81+far%C4%81%CA%BEi%E1%B8%8D+alqul%C5%ABb%E2%80%9D%3A+Using+automatic+transliteration+and+text+reuse+processes%2E+Master%E2%80%99s+thesis%2C+Haifa+Univ)
10. Aviad Rom. 2024. Processing dialectal Arabic with Transformerbased language models: Challenges and potential solutions. Thesis, Reichman University, Israel. G. M. Shahariar Shibli, Md. Tanvir Rouf Shawon, Anik Hassan Nibir, Md. [find](https://scholar.google.com/scholar?q=Aviad+Rom%2E+2024%2E+Processing+dialectal+Arabic+with+Transformerbased+language+models%3A+Challenges+and+potential+solutions%2E+Thesis%2C+Reichman+University%2C+Israel%2E+G%2E+M%2E+Shahariar+Shibli%2C+Md%2E+Tanvir+Rouf+Sha)
11. Zabed Miandad, and Nibir Chandra Mandal. 2023. Automatic back transliteration of Romanized Bengali (Banglish) to Bengali. Iran Journal of Computer Science, 6(1):69–80. ا ا ب ت ث ج ح خ د ذ ر ز س ش ص ض ط ظ ع غ ف ق ك ل م ن ه و ي ء ة ؤ ئ ى ε ب ت ث ج ح خ د ذ ر ز س ش ص ض ط ظ ع غ ف ق ك ل م ن ه و ي ء ة ؤ ئ ى ε 103 111 35 17 6142 33 97 847 3798 70 148 1989 1691 88 117 1017 1138 35 33 517 1 628 7 20 4 11329 18 3 2 53 46 *Table 5: Confusion matrix for all Arabic letters. Rows are predicted labels; columns are true labels.* Transliterated Arabic (output) [find](https://scholar.google.com/scholar?q=Zabed+Miandad%2C+and+Nibir+Chandra+Mandal%2E+2023%2E+Automatic+back+transliteration+of+Romanized+Bengali+%28Banglish%29+to+Bengali%2E+Iran+Journal+of+Computer+Science%2C+6%281%29%3A69%E2%80%9380%2E+%D8%A7+%D8%A7+%D8%A8+%D8%AA+%D8%AB+%D8%AC+%D8%AD+%D8%AE+%D8%AF+%D8%B0+%D8%B1+%D8%B2+%D8%B3+%D8%B4+%D8%B5+%D8%B6+)