Egyptian Arabic (arz) for machines

Egyptian Arabic is the spoken Arabic of Egypt and the dialect that the rest of the Arabic-speaking world knows best from film and television. The ISO 639-3 registration authority codes it arz, an individual living language inside the macrolanguage ara, Arabic, which has 30 member codes. Meta's NLLB-200 translation model and its FLORES-200 benchmark use the code arz_Arab and list the language as low resource. This post covers how Egyptian differs from Modern Standard Arabic (MSA) in ways a model can see, what public data exists, what the No Language Left Behind paper reports, what goes wrong, and how to test a system on it. The earlier post on why Arabic is hard for machines covers the script, vowel marks, and morphology that all varieties share; this one is about the Egyptian layer on top.

Egyptian Arabic vs MSA: what a model can see

Most descriptions of dialect difference start with sound, and some sound changes reach the written text. The MADAR corpus paper by Bouamor and colleagues gives the standard example: the letter qaf, which is /q/ in MSA and in Tunisian, is a glottal stop in Egyptian and Levantine and a /g/ in Gulf Arabic. Writers who spell by sound replace the qaf with a hamza; writers who spell by etymology keep it. The paper on Arabizi by Bies and colleagues lists two more Egyptian traits: the interdental sounds of MSA, written with the letters thaa and dhaal, became /t/ or /s/ and /d/ or /z/, so a word like thalj, ice, is pronounced talg; and the letter jeem, /j/ as in jam in MSA, is a hard /g/ in Egyptian, so jamal, camel, is gamal. The paper also notes that Egyptian has future prefixes ha- and Ha- where MSA uses sa-.

Vocabulary differs outright in common words. MADAR gives Egyptian bas, only, and tarabeza, table, against MSA faqaT and TAwila. Bies and colleagues give buSS, look, against MSA unZur. The function words matter most for a model, because they are the most frequent tokens in any sentence. Zaidan and Callison-Burch, in their 2014 Computational Linguistics article on dialect identification, show a Google Translate output of an Egyptian sentence in which the system could not consistently handle eh, what, elli, that, or da, this, which the paper counts among frequent words.

Negation is the grammatical difference that is easiest to spot in text. Zaidan and Callison-Burch note that the dialects allow circumfix negation, a particle before the verb and a suffix after it, and attached pronouns as indirect objects, both of which MSA lacks. In Egyptian the circumfix is ma- plus -sh. The CODA* guidelines by Habash and colleagues give Cairo examples: ma qalsh, he did not say; manish, I am not; mahiyyash, she is not. The same guidelines show Damascus using ma qaal with no suffix, so the suffix is one quick signal that separates Egyptian from northern Levantine text. MSA uses the standalone particles laa, lam, and lan with changed verb forms instead.

The MADAR paper, following Brustad, says syntactic differences between dialects are minor compared with morphology, and, following Benmamoun, that negation particles vary (ma, mish, muw, la, lam) while their distribution is largely uniform.

Spelling: the same word 27 ways

Egyptian has no official orthography, and the consequence for machines is variance rather than a single set of rules. The CODA* paper opens with a search experiment: 27 attested spellings of the Egyptian word for he does not say it, mabi'ulhaash, with frequencies from Google Search on September 29, 2017. The variance comes from independent choices: whether to attach the negation particle ma to the verb or write it separately, whether to write the stem by its sound with a hamza or by its origin with a qaf, and whether to spell the vowels long or short. The paper cites an earlier estimate that close to 24 percent of Egyptian Arabic words online are not spelled according to CODA, and that 32 percent of words in MSA comments online have spelling errors.

CODA, the Conventional Orthography for Dialectal Arabic, began in 2012 as an Egyptian convention and was extended in the 2018 paper to 28 city dialects, including Cairo, Alexandria, and Aswan. The convention prefers etymological spelling: write the qaf even where it is pronounced as a glottal stop. The MADAR CODA corpus paper by Eryani and colleagues gives the reason: any Arab speaker with minimal knowledge of MSA knows how that letter is pronounced in their own dialect, while a phonetic spelling with hamza might be read correctly by Egyptians but not by Gulf speakers. The 2018 paper's rule for the negation clitic is that ma is written as a separate word except when it attaches to a pronoun, as in manish. A system that normalizes Egyptian input before translation has to pick one of these conventions, and its training data almost certainly follows none of them consistently.

One script convention concerns the end of a word. The Unicode Arabic code chart defines U+0649 alef maksura as a yeh-shaped letter with no dots in any positional form, and U+064A yeh as the dotted letter. Bies and colleagues write that the two "are often used interchangeably in Egyptian Arabic writing conventions", and their annotators had to restore the distinction by hand. CODA* says the MSA rules for alef maksura apply, so a normalizer that follows CODA has to restore the dots from context. A search index that does not normalize the two code points will treat the Egyptian spelling and the MSA spelling of the same word as different strings, the same mismatch the post on why Arabic is hard describes for vowel marks.

Arabizi and code-switching

A large share of typed Egyptian is not in Arabic script at all. Arabizi is Arabic written in Latin letters with numerals for the letters that have no Latin equivalent. Bies and colleagues explain that the numeral 3 stands for ayn because it looks like a mirror image of the letter, and that the choice of Latin letter follows local pronunciation: j for jeem in Levantine Arabizi, g in Egyptian; sh in countries with an English second language, ch where French dominates. Darwish's 2013 paper describes Arabizi as common in informal settings such as social networks and often mixed with English, and reports 98.5 percent accuracy for detecting Arabizi words among English ones and 88.7 percent for converting them to Arabic script, with about a third of the conversion errors being spelling or morphological variants of the reference.

The scale of Arabizi in real Egyptian messaging is documented in the LDC's BOLT Egyptian Arabic SMS/Chat and Transliteration corpus, collected for DARPA's Broad Operational Language Translation program: 5,691 conversations, 262,026 messages, and 1,029,248 words of Egyptian text messages and chat, with about 1,856 Arabizi conversations transliterated into Arabic script and corrected by annotators. Bies and colleagues, describing a 1,503-conversation training set from the same collection effort, report that only 15 percent of conversations were entirely in Arabic script while 66 percent were entirely Arabizi, and that English words in Egyptian messages take Arabic morphology and determiners and are spelled the way an Egyptian pronounces them.

Code-switching into English is separate from Arabizi and happens in speech as well as text. The ArzEn corpus by Hamed, Vu, and Abdennadher is 12 hours of spontaneous Egyptian Arabic-English speech from 38 bilingual university students and employees, built for speech recognition and for studying how Egyptians switch. For a translation or search system the implication is the same: an Egyptian document may hold Arabic script, Latin-script Arabic, and English in one sentence, and a language identifier that assigns one label per sentence will be wrong about part of it.

What public data exists for Egyptian Arabic

Compared with other dialects, Egyptian is well served; compared with MSA, it is thin. The OPUS API lists six English-Egyptian corpora with about 173,900 aligned sentence pairs in total: wikimedia with 121,935, XLEnt with 22,047, WikiMatrix with 18,232, translatewiki with 10,937, Tatoeba with 740, and QED with 3. The same API for English-Arabic, under the generic code ar, lists 31 corpora and about 245 million pairs, led by OpenSubtitles with 87.9 million, CCMatrix and NLLB with 49.7 million each, and the United Nations corpus with 20.0 million. The ratio is roughly 1,400 to 1.

Almost all of the Egyptian Wikipedia-derived data traces to one source. The Egyptian Arabic Wikipedia's statistics page shows about 1.63 million content pages, 2.27 million total pages, 13.28 million edits, about 297,850 registered users, 283 active users, and 7 administrators. The List of Wikipedias ranks it 11th by article count, ahead of the Standard Arabic Wikipedia at 15th with about 1.34 million articles, even though the Standard Arabic edition has 5,544 active users to Egyptian's 283. Anyone training on it should look at what the articles contain. The Wikimedia site matrix shows that among Arabic varieties, only Arabic, Egyptian Arabic, and Moroccan Arabic have a Wikipedia at all; the Levantine, Iraqi, and Gulf codes have none.

The curated resources are smaller and better. The MADAR corpus has 2,000 sentences from the Basic Traveling Expression Corpus translated into all 25 of its city dialects, plus 10,000 more for five cities including Cairo, so Cairo has 12,000 sentences parallel with English, French, MSA, and the other four cities; Alexandria and Aswan have 2,000 each. The sentences average 6.5 words and are travel phrasebook material. The AraBench paper by Sajjad and colleagues describes the largest dialect-English training set, the Arabic-Dialect/English Parallel Text from Zbib and colleagues, with about 38,000 Egyptian-English sentences from weblogs, and pools it with MADAR and other sets into an evaluation suite. Speech data includes the LDC's CALLHOME Egyptian Arabic, 120 unscripted telephone conversations between native speakers of Egyptian Colloquial Arabic, in its Cairene variety, placed from North America, typically to Egypt. The NADI shared task data, described below, supplies Egyptian tweets with country labels.

What NLLB-200 reports for arz_Arab

The No Language Left Behind paper's language table lists arz_Arab, Egyptian Arabic, in the Semitic branch of Afro-Asiatic, as low resource, which the paper defines as fewer than 1 million publicly available de-duplicated sentence pairs with any other language. The paper includes eight Arabic varieties beyond MSA: Egyptian, Moroccan, Najdi, North Levantine, South Levantine, Ta'izzi-Adeni, Mesopotamian, and Tunisian. For the FLORES-200 test sets, the paper says the vendor teams analyzed how far each variety differed from MSA and then either translated from English or adapted the MSA set with the English as context. Footnote 7 says four were adapted, acm, acq, aeb, and ars, while Section 8.7.2 says the Egyptian and South Levantine sets were translated from English, and warns that any results involving them might be skewed by that different process.

The fairseq NLLB README links a metrics file for each released model. For arz_Arab paired with English, chrF++ is:

The 54.5B file also gives spBLEU with the SPM-200 tokenizer: 40.7 into English and 32.1 from English. For comparison, the same 3.3B file gives MSA, arb_Arab, 65.8 into English and 55.0 from English, so Egyptian trails MSA by about 7 chrF++ points into English and about 10 out of English. The 54.5B file also scores every dialect pair: Egyptian into MSA is 59.4 chrF++ and MSA into Egyptian 53.7, meaning the model does not treat them as the same language. The posts on NLLB-200 and FLORES-200 explain what the model sizes and metrics mean.

Section 8.7.2 of the paper is the most useful part for Egyptian. Table 43 measures how far each dialect reference in FLORES-200 devtest sits from the MSA reference: for Egyptian, 37.31 percent of tokens per sentence on average are not in the MSA sentence, 8.56 percent at the corpus level, and the Egyptian reference scores only 41.9 spBLEU against the MSA reference. Table 44 then measures the model's own English-to-Egyptian output against its English-to-MSA output: 28.33 percent sentence-level dialectness and 54.5 spBLEU. The output is closer to MSA than the human reference is. The paper reports that across the dialects the standard deviation of sentence-level dialectness fell from 14.6 in the human data to 8.8 in the generated data, a flattening of dialect differences. The paper also evaluated NLLB-200 on MADAR, mapping 16 of its 25 city dialects to the 8 codes; Table 58 gives 17.5 BLEU and 45.2 chrF++ from MSA into Egyptian and 27.4 and 49.5 from Egyptian into MSA.

What goes wrong: MSA out, dialect confusion, Arabizi in

The first failure is that a model answers in MSA when asked for Egyptian, or quietly treats Egyptian input as MSA. The NLLB dialectness numbers above are one measurement of the first. For the second, AraBench gives a clean experiment: a system trained on about 40 million MSA-English pairs scored 13.4 and 14.0 BLEU on Egyptian dev and test sets, better than a system trained on the 36,000 Egyptian training pairs alone, which scored 6.2 and 6.8, but far below the 18.7 and 19.3 reached by fine-tuning the MSA system on the Egyptian data. The paper credits the MSA system's showing to the lexical and orthographic similarities between MSA and the dialects. Zaidan and Callison-Burch's Figure 3 shows what the remaining failure looks like in output: a system that could not consistently handle the frequent words eh, elli, and da.

The second failure is dialect confusion in the step that decides what the input is. The NLLB paper's own language identifier was not scored on this decision: the caption of Table 5 says the Arabic languoids were merged after linguistic analysis when the identifier was evaluated on FLORES-200. The NADI 2020 shared task, with tweets from 100 provinces in 21 countries, had Egypt as its largest class, 6,635 tweets or 21.43 percent, and the best country-level system reached 26.78 macro F1. NADI 2024 moved to multi-label identification because many sentences are valid in several dialects, and the best system reached 50.57 macro F1. The paper reports that Egypt had the highest annotator agreement of the nine countries whose scores it reports, Fleiss kappa 0.69, and the highest geolocation precision, 94.6 percent, so Egyptian is the easy case, and even the easy case is far from solved. The paper also cites a finding that about 66 percent of a single-label classifier's errors were not true errors because the text was valid in both dialects.

The third failure is input in Latin letters: an Arabizi converter such as Darwish's gets 88.7 percent of words right, which leaves about one word in nine wrong before translation begins. The fourth is recognition of images. KITAB-Bench, a 2025 Arabic OCR benchmark of 8,809 samples across 9 domains, reports that the best end-to-end PDF-to-Markdown system reached 65 percent accuracy, that vision-language models beat traditional OCR tools by about 60 percent in character error rate, and that on line recognition Tesseract had a character error rate of 0.66 and EasyOCR 0.20. The benchmark's handwritten subset combines KHATT, ADAB, Muharaf, and OnlineKhatt, and the paper does not report results for Egyptian handwriting or chat screenshots. The post on OCR of scanned documents covers the recognition stage generally.

How large language models compare

The TARJAMAT paper by Kadaoui and colleagues tested ChatGPT, GPT-4, Bard, three commercial translators, and NLLB-200's distilled 1.3B model on 200 Egyptian sentences of at least ten words transcribed from television series, among ten Arabic varieties. NLLB run with MSA as the declared source scored 11.12 BLEU on Egyptian; run with Egyptian declared as the source it scored 13.87. ChatGPT scored 19.96 zero-shot and Google Translate 18.09. On MSA news sentences the order reversed: Google Translate 66.01, NLLB 41.34, ChatGPT 42.85. The gain from telling NLLB the source is Egyptian, about 2.8 BLEU, is the measured value of dialect routing on this test set.

How to evaluate an Egyptian Arabic system

• Test on Egyptian text, not MSA. FLORES-200 arz_Arab devtest is Wikinews, Wikijunior, and Wikivoyage text translated from English and is the public standard; MADAR Cairo, Alexandria, and Aswan give travel phrases; NADI gives tweets with country labels; the BOLT corpus gives SMS and chat with Arabizi.
• Report the direction. NLLB's own files put Egyptian into English about 14 chrF++ points above English into Egyptian on the 3.3B model.
• Measure dialectness of output, not only accuracy. The NLLB paper's method is to count tokens in the output that do not appear in the MSA translation of the same sentence; an Egyptian system whose output is 100 percent MSA tokens has not produced Egyptian.
• Include negation, the function words eh, elli, da, and the future prefix ha- in any spot check, because these are where MSA-trained systems break first.
• Test Arabizi and mixed-script input separately and report what fraction of words survive conversion.
• If documents are images, report character and word error rates on Egyptian samples of the same kind, since KITAB-Bench does not report results by dialect.

What to check

• Does the system name the variety as arz_Arab, Egyptian Arabic, or does it accept only ar or arb and treat everything as MSA?
• Which FLORES-200 score is quoted, which model size, and which direction, given that NLLB's files range from 55.6 to 60.8 chrF++ into English across sizes?
• Was the test set translated from English, like FLORES-200 arz_Arab, or taken from real Egyptian writing with its 27-way spelling variance?
• Does the pipeline normalize alef maksura and yeh, hamza and qaf, and attached and detached ma before matching or translating, and which convention does it follow?
• Does it detect and convert Arabizi, and what accuracy does it report on Egyptian SMS or chat?
• How much of the output is MSA when the requested target is Egyptian, measured against an MSA translation of the same input?

Azcuna's BABEL reads, translates, and searches foreign-language documents entirely on-device, with no cloud and no connection, and keeps every claim tied to the passage it came from.

Questions people ask

Does NLLB-200 support Egyptian Arabic?

Yes. The FLORES-200 README lists Egyptian Arabic under the code arz_Arab, and the No Language Left Behind paper's language table marks it as a low-resource language in Arabic script. The paper defines low resource as fewer than 1 million publicly available sentence pairs with any other language. The metrics files linked from the NLLB repository give FLORES-200 chrF++ scores for every released model, for example 58.6 into English and 44.8 from English for the 3.3B dense model.

What is arz_Arab?

arz_Arab is the NLLB-200 and FLORES-200 language code for Egyptian Arabic written in Arabic script. The first part is the ISO 639-3 code arz, which the ISO 639-3 registration authority defines as Egyptian Arabic, an individual living language and one of 30 members of the macrolanguage ara, Arabic. The second part is the ISO 15924 script code Arab. Modern Standard Arabic is a separate member of the same macrolanguage, coded arb.

What is the difference between Egyptian Arabic and MSA?

The MADAR corpus paper gives concrete cases. The letter qaf is a glottal stop in Egyptian where MSA has /q/. The future marker is ha- in Egyptian against sa- or sawfa in MSA. Common words differ outright: Egyptian bas for only and tarabeza for table against MSA faqaT and TAwila. The Bies et al. Arabizi paper adds that Egyptian lost the interdental sounds of MSA and pronounces jeem as a hard g. Negation wraps the verb with ma- and -sh.

How good is Egyptian Arabic to English translation?

On FLORES-200 devtest, the NLLB metrics files give chrF++ into English of 55.6 for the 600M distilled model, 58.4 for the 1.3B dense model, 58.6 for the 3.3B dense model, and 60.8 for the 54.5B mixture-of-experts model, whose spBLEU is 40.7. Modern Standard Arabic into English scores 65.8 chrF++ on the 3.3B model, about 7 points higher. The NLLB paper notes that the Egyptian test set was translated from English rather than adapted from MSA, which may skew comparisons.

Why does an Arabic translation model answer in MSA when the input is Egyptian?

Because almost all Arabic training text is Modern Standard Arabic. The AraBench paper reports about 40 million MSA-English sentence pairs against 38,000 Egyptian-English pairs in the largest public dialect corpus, and the OPUS API lists about 245 million English-Arabic pairs against about 174,000 English-Egyptian pairs. The No Language Left Behind paper also found a flattening of dialect differences in NLLB-200's output: the standard deviation of sentence-level dialectness fell from 14.6 in the human references to 8.8 in the model's translations.

What is Arabizi?

Arabizi is Arabic written in Latin letters, with numerals standing in for letters that have no Latin equivalent, such as 3 for the letter ayn. Kareem Darwish's 2013 paper describes it as common in informal settings and often mixed with English, and reports 98.5 percent accuracy detecting it and 88.7 percent converting it to Arabic script. In the Egyptian SMS and chat training set that Bies and colleagues described for DARPA's BOLT program, 66 percent of conversations were entirely Arabizi and only 15 percent entirely Arabic script.

How big is Egyptian Arabic Wikipedia?

The Egyptian Arabic Wikipedia's own statistics page shows about 1.63 million content pages, about 297,850 registered users, 283 active users, and 7 administrators. The List of Wikipedias on Wikimedia Meta ranks it 11th by article count, above the Standard Arabic Wikipedia's 1.34 million articles at 15th, although the Standard Arabic edition has 5,544 active users. The Wikimedia site matrix shows that among Arabic varieties only Arabic, Egyptian Arabic, and Moroccan Arabic have a Wikipedia.

Where can I get Egyptian Arabic parallel data?

The OPUS API lists six English-Egyptian corpora totaling about 173,900 aligned pairs: wikimedia with 121,935, XLEnt with 22,047, WikiMatrix with 18,232, translatewiki with 10,937, Tatoeba with 740, and QED with 3. The MADAR corpus has 12,000 Cairo sentences parallel with English, French, and MSA, plus 2,000 each for Alexandria and Aswan. The LDC distributes the BOLT Egyptian SMS and chat corpus and CALLHOME Egyptian Arabic speech.

How hard is it to tell Egyptian Arabic from other dialects?

Zaidan and Callison-Burch found crowd annotators 88.0 percent accurate at labeling dialect versus MSA on newspaper comments, with their classifier at 85.7 percent and an MSA-only system at 66.6 percent. In the NADI 2024 shared task, Egypt had the highest annotator agreement of any country, kappa 0.69, and the highest geolocation precision, 94.6 percent, yet the best overall country-level dialect identification system reached only 50.57 macro F1.

Sources

1. 1No Language Left Behind: Scaling Human-Centered Machine Translation · arXiv (NLLB Team, Meta AI), v3 of August 25, 2022
2. 2FLORES-200 (README) · Meta AI Research (facebookresearch on GitHub)
3. 3fairseq, NLLB branch: No Language Left Behind (README) · Meta AI Research (facebookresearch on GitHub)
4. 4NLLB-200 3.3B dense model: FLORES-200 metrics (metrics.csv) · Meta AI Research
5. 5NLLB-200 54.5B MoE model: FLORES-200 metrics (metrics.csv) · Meta AI Research
6. 6NLLB-200-Distilled 600M model: FLORES-200 metrics (metrics.csv) · Meta AI Research
7. 7NLLB-200 1.3B dense model: FLORES-200 metrics (metrics.csv) · Meta AI Research
8. 8ISO 639-3 code arz: Egyptian Arabic · SIL International, ISO 639-3 Registration Authority
9. 9ISO 639-3 code ara: Arabic (macrolanguage) · SIL International, ISO 639-3 Registration Authority
10. 10The MADAR Arabic Dialect Corpus and Lexicon · LREC 2018 (Bouamor et al.), European Language Resources Association
11. 11Unified Guidelines and Resources for Arabic Dialect Orthography · LREC 2018 (Habash et al.), European Language Resources Association
12. 12A Spelling Correction Corpus for Multiple Arabic Dialects · LREC 2020 (Eryani, Habash, Bouamor, Khalifa), European Language Resources Association
13. 13Arabic Dialect Identification · Computational Linguistics 40(1), 2014 (Zaidan and Callison-Burch), MIT Press
14. 14AraBench: Benchmarking Dialectal Arabic-English Machine Translation · COLING 2020 (Sajjad et al.), International Committee on Computational Linguistics
15. 15Transliteration of Arabizi into Arabic Orthography: Developing a Parallel Annotated Arabizi-Arabic Script SMS/Chat Corpus · EMNLP 2014 Workshop on Arabic Natural Language Processing (Bies et al.), Association for Computational Linguistics
16. 16Arabizi Detection and Conversion to Arabic · arXiv (Kareem Darwish), June 28, 2013
17. 17BOLT Egyptian Arabic SMS/Chat and Transliteration (LDC2017T07) · Linguistic Data Consortium, April 17, 2017
18. 18CALLHOME Egyptian Arabic Speech (LDC97S45) · Linguistic Data Consortium, 1997
19. 19ArzEn: A Speech Corpus for Code-switched Egyptian Arabic-English · LREC 2020 (Hamed, Vu, Abdennadher), European Language Resources Association
20. 20NADI 2024: The Fifth Nuanced Arabic Dialect Identification Shared Task · ArabicNLP 2024 (Abdul-Mageed et al.), Association for Computational Linguistics
21. 21NADI 2020: The First Nuanced Arabic Dialect Identification Shared Task · arXiv (Abdul-Mageed, Zhang, Bouamor, Habash), WANLP 2020
22. 22TARJAMAT: Evaluation of Bard and ChatGPT on Machine Translation of Ten Arabic Varieties · arXiv (Kadaoui et al.), v2 of October 23, 2023; ArabicNLP 2023
23. 23OPUS API: corpora for English-Egyptian Arabic (source=en, target=arz) · OPUS, the open parallel corpus (University of Helsinki / NLPL)
24. 24OPUS API: corpora for English-Arabic (source=en, target=ar) · OPUS, the open parallel corpus (University of Helsinki / NLPL)
25. 25Egyptian Arabic Wikipedia: Statistics (Special:Statistics) · Wikimedia Foundation
26. 26List of Wikipedias · Wikimedia Meta-Wiki
27. 27Wikimedia site matrix (API: action=sitematrix) · Wikimedia Foundation
28. 28Arabic, Range: 0600-06FF (code chart, The Unicode Standard) · Unicode, Inc.
29. 29KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding · arXiv (Heakl et al.), v2 of June 27, 2025; ACL 2025

Where this shows up in our work

• BABEL, offline multilingual translation and search translates and searches foreign-language documents on-device, every fact tied to its source passage.