Najdi and Gulf Arabic (ars) for machines

Gulf Arabic is the spoken Arabic of the Arabian Peninsula's eastern coast, and Najdi Arabic is the Arabic of the Najd region, which the dialect maps described below represent with Riyadh. The ISO 639-3 registration authority gives them separate codes, afb for Gulf Arabic and ars for Najdi Arabic, both individual living languages in the macrolanguage ara. Meta's NLLB-200 translation model and the FLORES-200 benchmark cover only one of them, Najdi, as ars_Arab, listed as low resource in the No Language Left Behind paper, and have no code for Gulf Arabic at all. This post covers what separates the peninsula dialects from Modern Standard Arabic (MSA) in text, why the Najdi FLORES-200 scores look so good, what public data exists for the Gulf, where systems fail on Emirati, Kuwaiti, and Qatari text, and how to evaluate. The post on why Arabic is hard for machines covers the script and morphology all varieties share.

Najdi, Gulf, and the codes that do and do not exist

The naming matters because the codes do not line up with ordinary usage. The Gumar corpus paper by Khalifa and colleagues, following Qafisheh and Holes, says that strictly speaking Gulf Arabic refers to the varieties spoken on the western coast of the Gulf, in Bahrain, Qatar, the seven Emirates, Kuwait, and the Al-Hasa region of eastern Saudi Arabia, and that Omani, Hijazi, Najdi, and Baharna Arabic are usually not included in grammars of Gulf Arabic because they vary considerably from that set. For their corpus the authors widened the term to any Arabic of the six Gulf Cooperation Council countries. The CODA* orthography guidelines draw the same lines on their dialect map: Riyadh under Najd, Jeddah under Hijaz, Doha, Abu Dhabi, and Manama under Gulf, and Muscat under Oman. The MADAR corpus is coarser and puts Doha, Muscat, Riyadh, and Jeddah in a single Gulf sub-region.

ISO 639-3 has a code for most of these: ars Najdi, afb Gulf, acw Hijazi, abv Baharna, acx Omani, adf Dhofari, ssh Shihhi, and for Yemen ayh Hadrami, ayn Sanaani, and acq Ta'izzi-Adeni. The FLORES-200 README lists two of them, ars_Arab, Najdi Arabic, and acq_Arab, Ta'izzi-Adeni Arabic. There is no afb_Arab. The NLLB paper's language table marks ars_Arab low resource, which the paper defines as fewer than 1 million publicly available de-duplicated sentence pairs with any other language, and names no specific variety for it, unlike the Baghdadi and Ammani it names for Iraqi and South Levantine. A program with text from Dubai, Kuwait City, or Doha therefore has three choices when it calls NLLB-200: declare the source as Najdi, as MSA, or as Ta'izzi-Adeni, none of which is the language of the text. The posts on NLLB-200 and FLORES-200 explain the model's codes and the benchmark.

Gulf Arabic vs MSA: what shows up in text

The Gumar paper lists the consonant changes that reach the spelling when writers spell by sound. The MSA letter qaf is pronounced /g/ or /j/, so qiddaam, in front, becomes jiddaam and qalam, pen, becomes galam. The letter kaf can be pronounced ch, so kabd, liver, becomes chebd; sheen can be pronounced ch, so shaay, tea, becomes chaay; and in some varieties jeem is pronounced y, so jalas, he sat, becomes yilas. The MADAR paper gives the same qaf fact from the other direction: the letter that is a glottal stop in Cairo and Beirut is a hard /g/ in Gulf Arabic. CODA*'s sound-to-letter table records Doha pronouncing shaaf, he saw, as chaaf. In CODA, all of these are written with the MSA letter, so a reader supplies the local sound; in text from forums and chat, writers choose.

Grammar differs in ways that are visible in the words themselves. Gumar reports that negation is marked by mu and its variants mub, muhub, and hub, that the future is marked by a ba- prefix instead of MSA sa-, and that in Kuwaiti, Bahraini, and some Saudi varieties the motion verb raaH has become a future marker. Possession can be marked by maal or Hag, so the book of the lady is li-ktaab maal il-mara. Case endings are lost, as in most dialects, with the exception of some Bedouin varieties. Some Saudi, Emirati, and Omani varieties keep a distinct feminine plural in verbs and pronouns, gaamat, she stood up, against gaaman, they stood up, which most other dialects have collapsed into the masculine. Zaidan and Callison-Burch, in their 2014 Computational Linguistics article, report the folk wisdom that Gulf is the closest regional dialect to MSA and note that it has preserved more of MSA's verb conjugation than other varieties, while cautioning that there are major differences. The MADAR paper's vocabulary comparison across 25 cities found Muscat the city dialect closest to MSA.

Spelling variance follows from the sound changes. The MADAR CODA corpus paper by Eryani and colleagues, which applied the CODA convention to 10,000 sentences from Beirut, Cairo, Doha, Rabat, and Tunis, gives the test case: the MSA word qamaruhu, his moon, written with qaf, is read correctly as amaru by an Egyptian and gamara by a Gulf speaker, while the phonetic spelling with hamza is readable by Egyptians but not Gulf speakers. A Gulf writer who spells galam with kaf, or with the Persian letter gaf that the NGA romanization table says is sometimes used outside Iraq for the g sound, produces a string no MSA dictionary holds. The Gumar authors found that an Egyptian-trained morphological analyzer had 5.6 percent out-of-vocabulary words on Gulf text against 9.3 percent for the MSA analyzer, which suggested working from the Egyptian model.

What public data exists for Gulf and Najdi Arabic

Parallel data with English is almost absent. The OPUS API for English-Najdi returns no corpora at all. The API for English-Gulf Arabic returns one, Tatoeba, with 81 aligned pairs. Neither code has a Wikipedia; the Wikimedia site matrix shows editions only for Arabic, Egyptian Arabic, and Moroccan Arabic. MADAR is the main parallel resource: 2,000 travel-phrase sentences translated into each of 25 city dialects, including Riyadh, Jeddah, Muscat, and Doha, plus 10,000 more for Doha, so Doha has 12,000 sentences parallel with English, French, MSA, and four other cities. The sentences average 6.5 words.

Monolingual Gulf text exists at far larger scale. Gumar holds 112,410,688 words in 9,335,224 lines from 1,236 documents, about 1,200 anonymous romance and drama novels published in online forums and collected by forum members into Word files. The authors labeled each document by country: Saudi 60.52 percent, Emirati 13.35 percent, Kuwaiti 5.91 percent, Omani 1.13 percent, Qatari 0.65 percent, Bahraini 0.49 percent, other Gulf 10.03 percent, and 7.93 percent MSA or other dialects. The novels are dialogue-heavy, written for a mainly female teenage audience, and the dialect of a novel is not necessarily the writer's own. The Annotated Gumar Corpus adds 200,110 tokens in 15,277 sentences from eight Emirati novels, annotated by hand for tokenization, part of speech, lemma, gloss, and spelling. The Bahrain Corpus by Abdulrahim and colleagues is 620,000 words of Bahraini folktales, comedy, plays, and cooking shows, with a 7,600-word hand-checked sample, built because Gumar, which the paper calls the largest Gulf Arabic corpus to date, holds a single genre of forum novels. CODA* lists Riyadh, Jeddah, Abu Dhabi, Manama, Doha, and Muscat among its 28 city dialects, so a spelling convention exists for each.

Two older sources carry Saudi text. Zaidan and Callison-Burch built the Arabic Online Commentary dataset from reader comments on three newspapers, one of them Al-Riyadh of Saudi Arabia, which supplied 805,000 comments and 18.8 million words, with about 40 percent of sentences dialectal across the set. The NADI 2020 shared task had 3,455 Saudi tweets from 10 provinces, 11.16 percent of its data and the third-largest class. The AraBench suite adds QAraC, a Qatari-English corpus of 14,700 sentences from TV series and talk shows, of which 6,713 form its test set, and notes that its dialectness level differs from MADAR's Doha set, so the same dialect looks more or less distant from MSA depending on the source.

What NLLB-200 reports for ars_Arab

The NLLB paper's footnote 7 says the Najdi FLORES-200 set was one of four Arabic sets adapted from the MSA translation rather than translated from English, with a term list of MSA words and Najdi equivalents. Table 43 measures the outcome against the MSA reference in devtest: 3.01 percent of tokens per sentence differ, 2.40 percent at the corpus level, and the Najdi reference scores 94.8 spBLEU and 96.5 chrF++ against the MSA reference. The paper's words are that "Najdi (ars_Arab) matches MSA almost exactly". By comparison the Egyptian reference differs in 37.31 percent of tokens and the Iraqi in 22.84 percent. Whatever the Najdi set measures, it is not the distance between Najdi speech and MSA.

The metrics files linked from the fairseq NLLB README give these chrF++ scores against English:

The 54.5B file adds spBLEU of 46.8 into English and 36.7 from English, against 48.3 and 43.0 for MSA. On the 3.3B model MSA scores 65.8 and 55.0, so Najdi trails by 2.7 chrF++ into English and 3.3 out, the smallest gap of any dialect adapted from MSA. The paper's Section 8.7.2 says performance is best for Najdi and explains it by how much the dialect benefits from transfer from the far larger MSA data. The 54.5B file also gives Najdi into MSA at 65.9 chrF++ and MSA into Najdi at 60.0, the highest of any dialect pair with MSA. Table 44 measures the model's English-to-Najdi output against its English-to-MSA output: 20.38 percent of tokens per sentence differ and the two score 62.8 spBLEU against each other. By that measure the output is more dialectal than the 3 percent reference.

On MADAR, where the paper mapped 16 of 25 city dialects to its 8 codes, Table 58 gives 19.8 BLEU and 45.5 chrF++ from MSA into Najdi and 27.8 and 49.3 from Najdi into MSA. The MADAR text is travel phrases. Najdi ranks second of the eight dialects there in both directions, behind North Levantine into the dialect and Ta'izzi-Adeni out of it, and its chrF++ sits well below the 60.0 and 65.9 that the 54.5B file gives for the same MSA pairs on FLORES-200.

What goes wrong: no code for the Gulf

The first failure is structural. TARJAMAT, by Kadaoui and colleagues, evaluated ChatGPT, GPT-4, Bard, three commercial translators, and NLLB-200's distilled 1.3B model on ten Arabic varieties, with 200 sentences of at least ten words per dialect transcribed from television series. For Emirati the authors could run NLLB only with MSA as the declared source, scoring 11.27 BLEU, and their table leaves the dialect-specific NLLB column empty for Emirati; NLLB-200 has no Emirati code. For Yemeni they mapped to Ta'izzi-Adeni and the score rose from 9.41 to 12.56; for Jordanian they mapped to South Levantine and it rose from 13.07 to 18.5. The Gulf has no such lever. ChatGPT scored 24.20 on the Emirati sentences zero-shot and Google Translate 19.56.

The second failure is dialect identification, which for the Gulf has a particular shape. NLLB's own language identifier was not scored on the dialect decision; the caption of its Table 5 says the Arabic languoids were merged after linguistic analysis when the identifier was evaluated. NADI 2024 recruited annotators from ten countries, none of them Saudi Arabia or the Emirates, and added 80 geolocated tweets from each of Lebanon, Libya, Saudi Arabia, and the UAE that were expected to be labeled invalid in the ten annotators' dialects. In its regional analysis Iraq is the only Gulf country, and the organizers say they plan to cover more of the Gulf. The best country-level system reached 50.57 macro F1 across the task. NADI 2020's best system reached 26.78 across 21 countries. Zaidan and Callison-Burch's note that Iraqi is sometimes folded into Gulf points the same way: an identifier may place a sentence in the right region and the wrong country, and for the peninsula the countries are where the vocabulary and the ch and y sounds differ.

The third failure is translating Gulf text into MSA so that MSA tools can run. The NADI 2023 and NADI 2024 shared tasks both included Emirati in a dialect-to-MSA translation subtask, with MADAR training data and newly curated test sentences whose source domain the organizers did not disclose. In 2023 the best open-track system scored 28.46 BLEU on Emirati, the highest of the four dialects that year; in 2024 the best system scored 23.38 on Emirati while the second-ranked system scored 12.53 on the same sentences, the widest gap between the top two systems for any dialect. The organizers concluded that dialect-to-MSA translation remains hard.

The fourth failure is recognition of images and informal text. KITAB-Bench, a 2025 Arabic OCR benchmark of 8,809 samples across 9 domains, reports that the best end-to-end PDF conversion reached 65 percent accuracy, that vision-language models beat traditional OCR tools by about 60 percent in character error rate, and that Tesseract's line-level character error rate was 0.66 against EasyOCR's 0.20. The handwriting samples come from KHATT, ADAB, Muharaf, and OnlineKhatt, and the paper does not report results by dialect. The post on OCR of scanned documents covers the recognition stage.

U.S. government materials

The National Geospatial-Intelligence Agency hosts the BGN/PCGN 1956 romanization system for Arabic, applied to geographic names in Bahrain, Kuwait, Oman, Qatar, Saudi Arabia, the United Arab Emirates, and Yemen, among other areas. The system is based "insofar as possible on fully pointed Modern Standard Arabic", so a Gulf place name pronounced with /g/ is romanized with q from its written qaf, and the notes say that recognizing dialectal deviations is essential and that colloquial forms found on authoritative sources should be kept. The post on transliterating names covers the consequences for matching. The Department of Education's ERIC index holds a Saudi Arabic Language and Culture Familiarization Course from 1966, a five-book programmed course of 66 everyday situations written at the University of Illinois with the Defense Language Institute and sponsored by the Office of Naval Research, designed to supplement the Defense Language Institute's course in Saudi Arabic. Current Department of State and Defense Language Institute pages on Gulf Arabic could not be read for this post, so no claim is made here about present designations.

How to evaluate a Gulf or Najdi Arabic system

• Decide which variety the documents are in before choosing a code. Riyadh text fits ars_Arab; Emirati, Kuwaiti, Qatari, and Bahraini text has no NLLB code, and a vendor who reports a Najdi score for it has measured something else.
• Do not accept the FLORES-200 ars_Arab score as a dialect score. The NLLB paper says the reference differs from MSA in 3 percent of tokens; compare the system's output on MADAR Doha, Riyadh, Jeddah, and Muscat, or on Gumar sentences, instead.
• Report direction and model size. The 3.3B model's Najdi scores are 63.1 into English and 51.7 out, and the 54.5B model is lower out of English than the 1.3B.
• Spot-check galam, chebd, chaay, yilas, the negators mu and mub, the future ba-, and maal and Hag possession, and compare the output with the MSA translation of the same sentence to see whether the model produced Gulf Arabic or MSA.
• Test identification on Saudi and Emirati text separately from Iraqi, since the MADAR corpus places Iraq and the Gulf cities in one Gulf region.
• If the source is scans or screenshots, report character and word error rates on Gulf samples, because KITAB-Bench does not report results by dialect.

What to check

• Does the system claim ars_Arab, Najdi Arabic, and what does it do with Emirati, Kuwaiti, or Qatari input, given that NLLB-200 has no code for afb?
• Which score is quoted, and was the quoting party aware that the NLLB paper says the Najdi test set matches MSA almost exactly?
• Does the test set contain the ch for kaf, y for jeem, and g or j for qaf spellings that Gulf writers use, or only MSA-adapted FLORES-200 text?
• Does the pipeline map those spellings back to the MSA letters before indexing, and is the same mapping applied to queries?
• Which country's text will the system read, and does the training data's country mix resemble it, given that Gumar is 60 percent Saudi?
• Is any claimed Gulf-to-English quality backed by a test set anyone can inspect, since the only public comparison, TARJAMAT, could not run NLLB in dialect mode on Emirati?

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

What is ars_Arab?

ars_Arab is the NLLB-200 and FLORES-200 code for Najdi Arabic, the dialect region that the CODA* guidelines represent with Riyadh, in Arabic script. The first part is the ISO 639-3 code ars, which the registration authority defines as Najdi Arabic, an individual living language in the macrolanguage ara, Arabic. The second part is the ISO 15924 script code Arab. The No Language Left Behind paper's language table lists it as low resource and names no specific variety.

Does NLLB-200 support Gulf Arabic?

Not under that name. ISO 639-3 codes Gulf Arabic as afb, an individual language, which the Gumar corpus paper places in Bahrain, Qatar, the Emirates, Kuwait, and eastern Saudi Arabia, but the FLORES-200 README and the NLLB paper's language table have no afb entry. The Arabian Peninsula varieties NLLB-200 does cover are ars_Arab, Najdi Arabic, and acq_Arab, Ta'izzi-Adeni Arabic from Yemen. A program with Emirati or Kuwaiti text has to pick one of those or Modern Standard Arabic as the declared source.

What is the difference between Najdi and Gulf Arabic?

The Gumar corpus paper, following Qafisheh and Holes, says Gulf Arabic strictly means the varieties of the western Gulf coast, in Bahrain, Qatar, the seven Emirates, Kuwait, and the Al-Hasa region of eastern Saudi Arabia, and that Omani, Hijazi, Najdi, and Baharna Arabic are usually left out of Gulf grammars because their features differ considerably. The CODA* guidelines place Riyadh under Najd and Doha, Abu Dhabi, and Manama under Gulf, while MADAR groups Doha, Muscat, Riyadh, and Jeddah in one Gulf sub-region.

Why is the Najdi FLORES-200 score so close to MSA?

Because the test set is nearly MSA. The No Language Left Behind paper says the Najdi set was adapted from the MSA translation rather than translated from English, and Table 43 shows the Najdi reference differs from the MSA reference in only 3.01 percent of tokens per sentence, with a reference-to-reference spBLEU of 94.8. The paper states that Najdi matches MSA almost exactly, and that the model's best results, on Najdi, can likely be explained by transfer from MSA.

How does Gulf Arabic differ from MSA in text?

The Gumar corpus paper lists the signals: the MSA letter qaf is pronounced g or j, as in galam for pen and jiddaam for front; kaf becomes ch, as in chebd for liver; sheen can become ch, as in chaay for tea; and jeem becomes y, as in yilas for he sat. Negation uses mu and its variants mub, muhub, and hub. The future is marked with ba- instead of MSA sa-, and in Kuwaiti and Bahraini the verb raaH has become a future marker. Possession uses maal or Hag.

Where can I get Gulf Arabic data?

The OPUS API lists no English-Najdi corpora and one English-Gulf corpus, Tatoeba, with 81 pairs. MADAR has 12,000 Doha sentences and 2,000 each for Riyadh, Jeddah, and Muscat, parallel with English, French, and MSA. Gumar is 112,410,688 words of Gulf forum novels with country labels, 60.52 percent Saudi and 13.35 percent Emirati; the Annotated Gumar Corpus adds 200,110 morphologically annotated Emirati tokens. The Bahrain Corpus has 620,000 words. Apart from MADAR and the 81 Tatoeba pairs, none of these has an English side.

How well does NLLB-200 translate Emirati Arabic?

There is no NLLB code for it, so the TARJAMAT paper, which evaluated ten Arabic varieties, could only run NLLB on its 200 Emirati sentences with MSA declared as the source and scored 11.27 BLEU, against 24.20 for ChatGPT and 19.56 for Google Translate. The paper's dialect-specific NLLB column is empty for Emirati. In the NADI 2023 and 2024 shared tasks, the best Emirati-to-MSA systems reached 28.46 and 23.38 BLEU.

Is Gulf Arabic closer to MSA than other dialects?

Zaidan and Callison-Burch report the folk wisdom that Gulf is the closest regional dialect to MSA and note that Gulf has preserved more of MSA's verb conjugation than other varieties, while cautioning that major differences remain. The MADAR paper found Muscat the city dialect closest to MSA by vocabulary overlap among its 25 cities. The NLLB measurements apply to Najdi only, and to a test set adapted from MSA, so they do not settle the question for Gulf Arabic.

What U.S. government material covers Gulf Arabic?

The Department of Education's ERIC index holds a 1966 Saudi Arabic Language and Culture Familiarization Course, a five-book programmed course of 66 situations sponsored by the Office of Naval Research and designed to supplement the Defense Language Institute's Saudi Arabic course. The National Geospatial-Intelligence Agency hosts the BGN/PCGN 1956 romanization system for Arabic place names in Bahrain, Kuwait, Oman, Qatar, Saudi Arabia, the United Arab Emirates, and Yemen, based on fully pointed MSA.

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 ars: Najdi Arabic · SIL International, ISO 639-3 Registration Authority
9. 9ISO 639-3 code afb: Gulf Arabic · SIL International, ISO 639-3 Registration Authority
10. 10ISO 639-3 code ara: Arabic (macrolanguage) · SIL International, ISO 639-3 Registration Authority
11. 11A Large Scale Corpus of Gulf Arabic · LREC 2016 (Khalifa, Habash, Abdulrahim, Hassan), European Language Resources Association
12. 12A Morphologically Annotated Corpus of Emirati Arabic · LREC 2018 (Khalifa et al.), European Language Resources Association
13. 13The Bahrain Corpus: A Multi-genre Corpus of Bahraini Arabic · LREC 2022 (Abdulrahim et al.), European Language Resources Association
14. 14The MADAR Arabic Dialect Corpus and Lexicon · LREC 2018 (Bouamor et al.), European Language Resources Association
15. 15Unified Guidelines and Resources for Arabic Dialect Orthography · LREC 2018 (Habash et al.), European Language Resources Association
16. 16A Spelling Correction Corpus for Multiple Arabic Dialects · LREC 2020 (Eryani, Habash, Bouamor, Khalifa), European Language Resources Association
17. 17Arabic Dialect Identification · Computational Linguistics 40(1), 2014 (Zaidan and Callison-Burch), MIT Press
18. 18AraBench: Benchmarking Dialectal Arabic-English Machine Translation · COLING 2020 (Sajjad et al.), International Committee on Computational Linguistics
19. 19NADI 2020: The First Nuanced Arabic Dialect Identification Shared Task · arXiv (Abdul-Mageed, Zhang, Bouamor, Habash), WANLP 2020
20. 20NADI 2023: The Fourth Nuanced Arabic Dialect Identification Shared Task · arXiv (Abdul-Mageed et al.), October 24, 2023; ArabicNLP 2023
21. 21NADI 2024: The Fifth Nuanced Arabic Dialect Identification Shared Task · ArabicNLP 2024 (Abdul-Mageed et al.), Association for Computational Linguistics
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-Najdi Arabic (source=en, target=ars) · OPUS, the open parallel corpus (University of Helsinki / NLPL)
24. 24OPUS API: corpora for English-Gulf Arabic (source=en, target=afb) · OPUS, the open parallel corpus (University of Helsinki / NLPL)
25. 25Wikimedia site matrix (API: action=sitematrix) · Wikimedia Foundation
26. 26Romanization of Arabic, BGN/PCGN 1956 System (Revised Presentation 2019) · U.S. Board on Geographic Names and Permanent Committee on Geographical Names, hosted by the National Geospatial-Intelligence Agency
27. 27Saudi Arabic Language and Culture Familiarization Course: Arab Cultural Assimilator, Books 1-5 (ERIC ED180190) · University of Illinois and Defense Language Institute, 1966, sponsored by the Office of Naval Research; indexed by ERIC, U.S. Department of Education
28. 28KITAB-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.