Arabic NER Annotation: Guidelines for Entities, Spans & Edge Cases

Arabic NER annotation means marking the names of people, places, organizations and other entities in Arabic text, with a type and exact span boundaries. It is harder than in English for three reasons: Arabic has no capital letters, conjunctions and prepositions attach directly to names, and many given names are also ordinary adjectives. A usable Arabic NER guideline therefore fixes span rules for clitics, the article and titles, defines how to handle names that double as common words, and picks a published entity scheme before anyone labels.

Below: span rules with worked examples, Arabic name structure, sector taxonomies and a dataset table checked against the papers. For tokenization, normalization and tool setup, see our playbook on how to annotate Arabic text.

Why is Arabic named entity recognition hard?

Khaled Shaalan's survey of Arabic NER (Computational Linguistics, 2014) lists the causes. Four shape the guideline:

• No capitalization. Arabic script has none, and the survey notes that most Arabic proper nouns look exactly like common nouns or adjectives.
• Agglutination. Conjunctions (و، ف) and prepositions (ب، ل، ك) attach to the name, so one entity has many surface strings.
• Missing short vowels. Unvoweled قطر can be Qaṭar, the country, or quṭr, "diameter"; only context decides.
• Spelling variation. Foreign names are transliterated many ways; the survey's example is "Washington".

Leave any of these undecided and every annotator decides differently.

Arabic NER span rules: clitics, the article, titles and nesting

Two conventions are defensible. Word-level spans cover whole whitespace tokens. Wojood (Jarrar et al., LREC 2022) works this way: attached prefixes and suffixes count as part of the entity name. Clitic-excluded character spans start after attached conjunctions and prepositions. The difference is portability: clitic-excluded spans can always be widened to whole words, while word-level spans need a morphological segmenter to narrow them. This guide recommends clitic-excluded character spans, with the definite article ال kept inside the span because it is part of the name as written.

The title, nesting and event rows follow Wojood's guidelines: person names exclude titles (except prophets, kings and similar), occupations get their own OCC type, and the Bank of Cairo manager is the paper's own nesting example. The descriptor row follows CLEANANERCorp (Al-Duwais, Al-Khalifa and Al-Salman, OSACT 2024), whose MUC-based re-annotation of ANERcorp includes words such as منظمة and شركة in organization names. The Arabic annotation guidelines template has a slot for each of these rules.

Names that are also ordinary words

Shaalan's classic case is أشرف: the name Ashraf, the verb ashrafa, "he supervised", or the superlative ashraf, "most honorable". Many more are everyday words:

Four rules:

1. Tag function, not form. A word is an entity only when it names a specific referent here. A gazetteer hit is a hint, not a label.
2. Write the cue words down. Titles (السيد، الشيخ، الأستاذة), kinship particles (بن، أبو) and place cues (مسقط رأسه) do the work capital letters do in English; Shaalan calls them trigger words.
3. Allow an "ambiguous" flag rather than forcing a guess. CLEANANERCorp reports that the first words of many ANERcorp sentences had no clear correct label.
4. Seed gold items with homonyms, the same word as a name and as an adjective, to measure this confusion directly.

How Arabic names are structured, and where the span ends

The IFLA guide to Arabic personal names (2018) sets out the traditional name elements used below; the example names are common, widely attested ones.

The nisba row hides a common confusion. In الوفد السعودي (al-wafd as-Saʿūdī, "the Saudi delegation"), السعودي is a nationality adjective. It is not a surname and belongs in NORP if your scheme follows OntoNotes. Modern names rarely carry all five elements, and usage differs by country, so take guideline examples from your own data, not classical texts.

Entity taxonomies for finance, government and healthcare

The AMWAL authors (Abdo, Hatekar and Cavar, FinNLP 2025) note that most Arabic NER systems remain generic, medicine aside. Start from a published base; add only what your product uses:

Domain types need domain annotators: a banker knows a sukuk is an instrument; a physician knows which abbreviation is a drug.

Arabic named entity recognition datasets

Figures are as reported by each paper or dataset page (as of September 2026):

Two cautions. ANERcorp scores across papers are hard to compare: the CAMeL Lab page explains that copies drifted between users until the agreed 2020 split. And none of these corpora is your domain. Use public data to pre-train and choose a scheme; build the test set from the text your model will see, in every dialect you serve (see why dialect coverage sets your model's ceiling).

How do you QA Arabic NER annotation?

Wojood and CLEANANERCorp both document reusable quality processes.

• Screen on hard items. Wojood gave about 30 trained candidates a deliberately hard 750-token quiz and recruited the 12 who scored above 90%.
• Review by entity list, not only by sentence. Wojood's experts reviewed a list of every annotated entity, which exposes one string carrying different labels. CLEANANERCorp found exactly that in ANERcorp: الضفة الغربية (aḍ-Ḍiffa al-Gharbiyya, "the West Bank") was tagged LOC in some sentences and O in others.
• Let a model flag, and people decide. Wojood retrained a tagger on its own labels and had annotators check every mismatch, twice. CLEANANERCorp ran confident learning (the cleanlab library), which flagged 1,945 issues for manual review.
• Measure agreement the NER way. The Wojood paper notes that Cohen's kappa can overstate NER agreement because most tokens are O. Report pairwise span-level F1 per type, and separate boundary disagreements from type disagreements; they have different fixes.
• Test pre-labels before trusting them. In LabelBench, a different task, a local language model scored 78% on binary Arabic toxicity. On the real multi-label taxonomy, its exact-match accuracy on the fixed held-out manifest fell to 25.3%. Measure any pre-labeler on your own gold first.

Bayanat Labs follows a Vet, Train, Produce, QA, Deliver method. Contributors, including licensed professionals for financial, legal and medical text, pass dialect and domain screening before touching client data; work is calibrated against gold standards with adjudication, rework and an audit trail. Our Arabic data annotation services include NER and span labeling across 25+ Arabic varieties, hosted in-region by default.