---
title: "Romanized Arabic Transliteration — Chalabi 2012"
description: "The authors crawled blogs to build a monolingual colloquial corpus and had a 35K-word… 1% accuracy. 13 references, cited by 15 papers here."
authors: ["Achraf Chalabi", "Hany Gerges"]
year: 2012
venue: "Proceedings of the Second Workshop on Advances in Text Input Methods"
source: "https://aclanthology.org/W12-4808/"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/chalabi-2012-romanized-arabic-transliteration/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/chalabi-2012-romanized-arabic-transliteration/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2012](https://md.research.elamur.ai/free/en/papers/2012/page/2/) / Romanized Arabic Transliteration

# Romanized Arabic Transliteration

**Authors:** Achraf Chalabi and Hany Gerges · **Year:** 2012 · **Venue:** Proceedings of the Second Workshop on Advances in Text Input Methods · **Publisher:** [International Committee on Computational Linguistics (ICCL)](https://md.research.elamur.ai/free/en/publishers/international-committee-on-computational-linguistics-iccl/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/W12-4808/) · [PDF](https://aclanthology.org/W12-4808.pdf) · [OpenAlex](https://openalex.org/W2251092354)

**Citations:** 23 citations (OpenAlex, 2026-10-14) · [15](https://md.research.elamur.ai/free/en/papers/chalabi-2012-romanized-arabic-transliteration/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper addresses converting romanized Arabic (Arabizi, Franco Arabic), the Latin-letter way many Arabs type colloquial Arabic, into Arabic script, which is easier to read. Two obstacles stand in the way: there is no fixed spelling convention, so one Arabic form maps to many romanizations, and there is no ready parallel data. The authors crawled blogs to build a monolingual colloquial corpus and had a 35K-word romanized vocabulary translated by hand. Their engine is hybrid. Mapping rules, either written by linguists or learned through alignment, generate candidate spellings. These candidates are then ranked with a log-linear mix of a transliteration model, a character trigram model and a word unigram model, with weights tuned for MRR. Pruning and affix stripping keep the response real-time. On a 5K-word test set, the top candidate is correct 85% of the time. The engine runs as a Windows TSF input service and also handles out-of-vocabulary names in MT. It was extended to Hebrew and Farsi by mining name pairs from MT parallel sentences, which makes the framework cheap to port to new scripts.

## Key points
- The data came from the in-house "Blog Muncher" crawler, which built the colloquial Arabic corpus. A 35K-word parallel romanized-to-Arabic list was translated manually to train the transliteration model.
- The score is λT·log PT(s|t) + λC·log PC(t) + λW·log PW(t):
  - PC is a KN-smoothed character trigram model.
  - PW is a word unigram model.
  - Both language models were built with the MSR LM toolkit.
  - 10% of the parallel data was used for weight tuning and another 10% was held out for testing.
- Pruning happens at each character position. The default pruning factor was chosen to cost no more than 1% accuracy. With a 1300-rule table, the loss was only 0.1% at 100-best. A time limit triggers more aggressive pruning, down to first-best if needed.
- Results on 5K test words were 85% top-candidate accuracy, 95% linear scoring and 90% MRR over the 5-best list.
- The system was ported to Hebrew-to-English and Farsi-to-English using parallel names extracted from MT corpora.

## Interesting ideas
- Hand-written or alignment-learned rules only propose candidates, and separate statistical models decide among them, so expert knowledge and data can be swapped depending on what a language offers.
- The pruning threshold was picked by measuring accuracy loss against beam size and accepting at most 1% loss, plus a time-triggered fallback that narrows the beam progressively.
- When a word yields no known candidate, the system strips listed affixes, transliterates the stem, and reattaches the affixes' known forms only if the stem is attested in the target lexicon.
- A crude letter mapping bootstraps a weak transliterator that is run over parallel sentences. Name pairs are spotted by edit distance plus cosine similarity, and the rule set is then grown iteratively via monotonic alignment.
- The system is scored both on top-1 accuracy and on MRR over an n-best list, matching whether the downstream consumer can choose among options or not.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (tokens): Lemma and morphology guesses for unknown or noisy words could come from explicit rules and be ranked by a character/word LM, keeping every proposal traceable to a named rule.
- **I2** (tree): The beam in the perceptron parser or gparse decoding could be calibrated the same way, with a documented accuracy-versus-speed curve and a deadline-driven degradation mode.
- **I3** (tokens): Out-of-vocabulary handling could try prefix/suffix stripping and accept the analysis only when the stem exists in the lexicon, giving an explainable fallback instead of a blind guess.
- **I4** (learn): Silver data could be grown the same way, letting a weak seed model propose matches, keeping only those confirmed by an independent similarity check, and retraining in rounds.
- **I5** (answer): Readers could report both strict first-answer accuracy and rank-aware scores over candidate answers, showing whether errors come from ranking or from missing candidates.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Microsoft](https://md.research.elamur.ai/free/en/companies/microsoft/) — Company, Redmond, United States · 378 works · 161,630 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works

## Cite

Cite as: Achraf Chalabi and Hany Gerges. 2012. Romanized Arabic Transliteration. Proceedings of the Second Workshop on Advances in Text Input Methods.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{chalabi2012romanized,
  title = {{Romanized Arabic Transliteration}},
  author = {Achraf Chalabi and Hany Gerges},
  booktitle = {Proceedings of the Second Workshop on Advances in Text Input Methods},
  year = {2012},
  url = {https://aclanthology.org/W12-4808/}
}
```

- [← previous A Fully Coreference-annotated Corpus of Scholarly Papers from the ACL Anthology](https://md.research.elamur.ai/free/en/papers/schafer-2012-fully-coreference-annotated-corpus-scholarly/)
- [next → The Rocky Road towards a Swedish FrameNet - Creating SweFN](https://md.research.elamur.ai/free/en/papers/heppin-2012-rocky-road-swedish-framenet-creating/)
- [↑ back to papers of 2012](https://md.research.elamur.ai/free/en/papers/2012/page/2/)

## References

13 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Yaser Al-Onaizan and Kevin Knight. 2002. Machine transliteration of names in Arabic text. In ACL Workshop on Comp. Approaches to Semitic Languages. [find](https://scholar.google.com/scholar?q=Yaser+Al%2DOnaizan+and+Kevin+Knight%2E+2002%2E+Machine+transliteration+of+names+in+Arabic+text%2E+In+ACL+Workshop+on+Comp%2E+Approaches+to+Semitic+Languages%2E)
2. Shane Bergsma and Grzegorz Kondrak. 2007. Alignment-based discriminative string similarity. In ACL, pages 656–663, Prague, Czech Republic, June. [find](https://scholar.google.com/scholar?q=Shane+Bergsma+and+Grzegorz+Kondrak%2E+2007%2E+Alignment%2Dbased+discriminative+string+similarity%2E+In+ACL%2C+pages+656%E2%80%93663%2C+Prague%2C+Czech+Republic%2C+June%2E)
3. Slaven Bilac and Hozumi Tanaka. 2004. A hybrid back-transliteration system for Japanese. In COLING, pages 597–603, Geneva, Switzerland. [find](https://scholar.google.com/scholar?q=Slaven+Bilac+and+Hozumi+Tanaka%2E+2004%2E+A+hybrid+back%2Dtransliteration+system+for+Japanese%2E+In+COLING%2C+pages+597%E2%80%93603%2C+Geneva%2C+Switzerland%2E)
4. Dayne Freitag and Shahram Khadivi. 2007. A sequence alignment model based on the averaged perceptron. In EMNLP, pages 238–247, Prague, Czech Republic, June. [find](https://scholar.google.com/scholar?q=Dayne+Freitag+and+Shahram+Khadivi%2E+2007%2E+A+sequence+alignment+model+based+on+the+averaged+perceptron%2E+In+EMNLP%2C+pages+238%E2%80%93247%2C+Prague%2C+Czech+Republic%2C+June%2E)
5. Ulf Hermjakob, Kevin Knight, and Hal Daum´e III. 2008. Name translation in statistical machine translation - learning when to transliterate. In ACL, pages 389–397, Columbus, Ohio, June. [find](https://scholar.google.com/scholar?q=Ulf+Hermjakob%2C+Kevin+Knight%2C+and+Hal+Daum%C2%B4e+III%2E+2008%2E+Name+translation+in+statistical+machine+translation+%2D+learning+when+to+transliterate%2E+In+ACL%2C+pages+389%E2%80%93397%2C+Columbus%2C+Ohio%2C+June%2E)
6. Sittichai Jiampojamarn, Colin Cherry, and Grzegorz Kondrak. 2008. Joint processing and discriminative training for letter-to-phoneme conversion. In ACL, pages 905–913, Columbus, Ohio, June. [find](https://scholar.google.com/scholar?q=Sittichai+Jiampojamarn%2C+Colin+Cherry%2C+and+Grzegorz+Kondrak%2E+2008%2E+Joint+processing+and+discriminative+training+for+letter%2Dto%2Dphoneme+conversion%2E+In+ACL%2C+pages+905%E2%80%93913%2C+Columbus%2C+Ohio%2C+June%2E)
7. Kevin Knight and Jonathan Graehl. 1998. [Machine transliteration](https://md.research.elamur.ai/free/en/papers/knight-1998-machine-transliteration/). Computational Linguistics, 24(4):599–612.
8. Colin Cherry and Hisami Suzuki. 2009. Discriminative Substring Decoding for Transliteration, In EMNLP. [find](https://scholar.google.com/scholar?q=Colin+Cherry+and+Hisami+Suzuki%2E+2009%2E+Discriminative+Substring+Decoding+for+Transliteration%2C+In+EMNLP%2E)
9. Philipp Koehn, Franz J. Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In HLT-NAACL. [↗](https://doi.org/10.21236/ada461156)
10. Haizhou Li, Min Zhang, and Jian Su. 2004. [A joint source-channel model for machine transliteration](https://md.research.elamur.ai/free/en/papers/li-2004-joint-source-channel-model-machine/). In ACL, pages 159–166, Barcelona, Spain, July.
11. Eric Sven Ristad and Peter N. Yianilos. 1998. Learning string-edit distance. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(5):522–532. [↗](https://doi.org/10.1109/34.682181)
12. Tarek Sherif and Grzegorz Kondrak. 2007b. Substring-based transliteration. In ACL, pages 944– 951, Prague, Czech Republic, June. [find](https://scholar.google.com/scholar?q=Tarek+Sherif+and+Grzegorz+Kondrak%2E+2007b%2E+Substring%2Dbased+transliteration%2E+In+ACL%2C+pages+944%E2%80%93+951%2C+Prague%2C+Czech+Republic%2C+June%2E)
13. Reinhard Kneser and Hermann Ney. 1995. Improved backing-off for m-gram language modeling. In International Conference on Acoustics, Speech, and Signal Processing (ICASSP-95), pages 181–184. [find](https://scholar.google.com/scholar?q=Reinhard+Kneser+and+Hermann+Ney%2E+1995%2E+Improved+backing%2Doff+for+m%2Dgram+language+modeling%2E+In+International+Conference+on+Acoustics%2C+Speech%2C+and+Signal+Processing+%28ICASSP%2D95%29%2C+pages+181%E2%80%93184%2E)

*Updated 1 November 2026*