By skyforbes 
A systematic survey and research agenda on Historical perspectives, datasets, benchmarks, Models, Tokenization, Morphology, and Evaluation in Arabic NLP
tl;dr
- Arabic NLP evolved from rule‑based analyzers (1980s–90s) → statistical ML (2000s) → neural/transformers and LLMs (2020s)
- Core hurdles: diglossia, dialects, rich morphology & clitics, undiacritized orthography, code‑switching, RTL/Unicode.
- Models: AraBERT/ArabicBERT, MARBERT (dialects), XLM‑R/mT5 (multilingual), Noor (10B), Jais (13B).
- Benchmarks & tasks: ALUE/ARLUE, NADI (dialect ID), NER, QA, MT, summarization, toxicity, diacritization.
- Data bedrock: PATB, OSCAR, Gigaword, MADAR, Tashkeela, social‑media corpora; urgent gaps in Maghrebi/Gulf/Sudanese dialects and code‑switch.
- What to do next: morphology‑aware tokenization, dialect expansion, open, reproducible eval, bias & safety audits, ethically‑governed datasets.
The Arabic internet and enterprise content are exploding across MENA and the diaspora. High‑impact use cases — search & retrieval, customer care, moderation & risk, knowledge…
Learn more about A Comprehensive Review of Arabic NLP — From Calligraphy to Transformers