---
title: "Semitic languages — papers — Elamur Research"
description: "Semitic languages: 1,485 papers of the Elamur Research library, the most cited first. Among them: AraBERT: Transformer-based Model for…"
canonical: "https://research.elamur.ai/free/en/topics/semitic-languages/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/semitic-languages/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Semitic languages

# Semitic languages

1,485 papers · page 1 of 16 · 1–25 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

Arabic, Hebrew, Aramaic, Amharic and their kin: root-and-pattern words, scripts without vowels.

A survey · about 19 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What it is](#a-what-it-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

The Semitic languages form the largest living branch of the Afroasiatic family. Arabic, with its standard written form and dozens of spoken dialects, is by far the most widely used; Hebrew, revived as a spoken language in the late nineteenth and early twentieth centuries, Amharic and Tigrinya in the Horn of Africa, Maltese, and the Neo-Aramaic varieties of the Middle East are its other living members. Behind them stand a long line of ancient languages known from inscriptions, clay tablets and scripture: Akkadian, Ugaritic, Phoenician, Biblical Hebrew, the Aramaic of the Bible and the Talmud, Syriac and Ge'ez. What unites them is a shared ancestry and a strikingly similar way of building words: most content words are formed by weaving a consonantal root, often three consonants long, into a pattern of vowels and affixes.

That root-and-pattern structure, together with writing systems that usually leave short vowels out, shapes almost every problem studied on this topic. A single unvowelled string of Arabic or Hebrew letters may stand for a noun, a verb or a phrase with attached prepositions, conjunctions and pronouns, and the reader resolves the ambiguity from context. Computers must do the same. As a result, research on Semitic languages has long been dominated by morphological analysis, segmentation and diacritization, and only later broadened into parsing, translation, dialect identification, sentiment and offensive-language detection, speech recognition and, most recently, large language models built specifically for Arabic and Hebrew.

The papers here come from two traditions that rarely cite each other. One is computational: Arabic and Hebrew natural language processing, from finite-state morphology in the 1990s to pretrained transformers after 2019. The other is philological and theoretical linguistics: the grammar of Biblical Hebrew, Masoretic reading traditions, Neo-Aramaic dialect syntax, Ethiopian Semitic phonology. This article covers both, weighted towards the computational side.

## What it is

### The family and its varieties

Semitic languages are usually divided into an East branch, represented by the extinct Akkadian of Mesopotamia, and a West branch that includes Central Semitic (Arabic, the Canaanite languages such as Hebrew and Phoenician, Aramaic, Ugaritic) and South Semitic (the Ethiopian languages and the Modern South Arabian languages). For computational work the practical division is different. Arabic is studied in three broad varieties: Classical Arabic, the language of the Qur'an and of pre-modern literature; Modern Standard Arabic (MSA), the shared written standard of news, education and formal speech; and the regional dialects (Egyptian, Levantine, Gulf, Maghrebi, Iraqi and many more), which people actually speak and increasingly write online, sometimes in Latin letters and digits ("Arabizi"). Hebrew is likewise split between Modern Hebrew and the older layers of Biblical, Mishnaic and Rabbinic Hebrew, each with its own corpora and research community.

### Core notions

- **Root and pattern.** A root such as Arabic k-t-b ("writing") combines with templates to give *kataba* "he wrote", *kitāb* "book", *maktab* "office", *kātib* "writer". Hebrew works the same way through its *binyanim* (verb patterns) and noun patterns. Some patterns are inflectional: the Arabic "broken plural" changes the internal vowels of a noun rather than adding a suffix.
- **Non-concatenative morphology.** Because roots and patterns interleave, the classical model of morphology as a chain of morphemes does not fit. Computational linguists answered with multi-tier and multi-tape finite-state models, discussed below.
- **Clitics and tokenization.** Conjunctions (*wa-* "and"), prepositions (*bi-*, *li-*), the definite article (*al-*, Hebrew *ha-*) and possessive or object pronouns are written attached to the word. What counts as a "token" is therefore itself a design decision.
- **Abjad scripts and diacritics.** Arabic and Hebrew scripts mark consonants and long vowels, while short vowels, consonant doubling and Arabic case endings are optional marks (diacritics in Arabic, *niqqud* in Hebrew) that are mostly omitted in everyday text. Restoring them is a task in its own right.
- **Diglossia.** In the Arab world the written standard and the spoken dialects differ in vocabulary, morphology and syntax, and the dialects have no agreed spelling. Tools trained on MSA degrade sharply on dialect.

### How it differs from neighbouring topics

The topic overlaps heavily with [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/), which concentrates on the Northwest Semitic languages and their ancient texts, and with [Morphology](https://md.research.elamur.ai/free/en/topics/morphology/), for which Semitic languages are a classic test case. It differs from [Akkadian and cuneiform NLP](https://md.research.elamur.ai/free/en/topics/akkadian-cuneiform-nlp/) mainly in scope: cuneiform work is a small, specialised corner of the same family. Unlike [Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/) in general, Arabic is spoken by hundreds of millions of people; its difficulty lies less in the amount of text than in the gap between the written standard and the dialects, and in the morphology.

## History and milestones

### Linguistic foundations

The comparative study of Semitic languages is among the oldest branches of linguistics. Medieval Arab grammarians such as Sibawayh described Arabic in terms of roots and measures, and Jewish grammarians of the same period applied similar ideas to Hebrew. Nineteenth- and twentieth-century philology established the family tree and reconstructed Proto-Semitic. That tradition continues in this library through studies such as [The Tiberian Pronunciation Tradition of Biblical Hebrew](https://md.research.elamur.ai/free/en/papers/khan-2020-tiberian-pronunciation-tradition-biblical/), [Studies in Semitic Vocalisation and Reading Traditions](https://md.research.elamur.ai/free/en/papers/hornkohl-2020-studies-semitic-vocalisation-reading/) and [sound changes in the reading tradition of Biblical Aramaic](https://md.research.elamur.ai/free/en/papers/suchard-2019-sound-changes-pre-masoretic-reading/).

In theoretical linguistics, John McCarthy's autosegmental analysis of Arabic verb morphology around 1980 treated the root, the vowel melody and the consonant-vowel skeleton as separate tiers. That idea fed directly into computational morphology and still informs work such as [a templatic account of Tigre vowel harmony](https://md.research.elamur.ai/free/en/papers/faust-2017-get-that-into-your-head/) and the recent [Pattern-and-root inflectional morphology: the Arabic broken plural](https://md.research.elamur.ai/free/en/papers/neme-2026-pattern-root-inflectional-morphology-arabic/).

### Finite-state morphology (1980s–2000s)

Early computational work tackled the interleaving of root and pattern with finite-state machinery. George Kiraz modelled the separate tiers as separate tapes of a multi-tape automaton, in [Multi-Tape Two-Level Morphology](https://md.research.elamur.ai/free/en/papers/kiraz-1994-multi-tape-two-level-morphology/) and later in [multitape automata for Syriac and Arabic](https://md.research.elamur.ai/free/en/papers/kiraz-2000-multitiered-nonlinear-morphology-multitape-finite/). Kenneth Beesley at Xerox showed that standard finite-state operations suffice if roots and patterns are merged by a compile-time step, in [Arabic Morphology Using Only Finite-State Operations](https://md.research.elamur.ai/free/en/papers/beesley-1998-arabic-morphology-only-finite-state/). For Hebrew, [A Finite-State Morphological Grammar of Hebrew](https://md.research.elamur.ai/free/en/papers/yona-2005-finite-state-morphological-grammar-hebrew/) played a similar role.

Tim Buckwalter's Arabic morphological analyzer, released through the Linguistic Data Consortium (LDC) in the early 2000s, took a more pragmatic route: large tables of prefixes, stems and suffixes with compatibility rules. It became the backbone of Arabic NLP for two decades, and its practical difficulties are recorded in [Issues in Arabic Orthography and Morphology Analysis](https://md.research.elamur.ai/free/en/papers/buckwalter-2004-issues-arabic-orthography-morphology-analysis/).

### Treebanks and statistical disambiguation (2002–2012)

A dedicated workshop series, Computational Approaches to Semitic Languages, began in 2002 and gave the field a venue. Its first meeting already contained [a shallow Arabic analyser built in a day](https://md.research.elamur.ai/free/en/papers/darwish-2002-building-shallow-arabic-morphological-analyser/) and [Gal's HMM vowel restoration for Arabic and Hebrew](https://md.research.elamur.ai/free/en/papers/gal-2002-hmm-approach-vowel-restoration-arabic/). At the same time the LDC built the Penn Arabic Treebank, described in [Maamouri and Bies's account of building it](https://md.research.elamur.ai/free/en/papers/maamouri-2004-developing-arabic-treebank-methods-guidelines/). Annotated text made supervised learning possible, and the "analyze, then choose" design of [Habash and Rambow's "One Fell Swoop" paper](https://md.research.elamur.ai/free/en/papers/habash-2005-arabic-tokenization-part-speech-tagging/) set the pattern for the decade.

Hebrew research followed a parallel path at Israeli universities, with morpheme-level models such as [an unsupervised morpheme-level HMM for Hebrew](https://md.research.elamur.ai/free/en/papers/adler-2006-unsupervised-morpheme-based-hmm-hebrew/) and joint models of segmentation and syntax such as [integrated morphological and syntactic disambiguation of Hebrew](https://md.research.elamur.ai/free/en/papers/tsarfaty-2006-integrated-morphological-syntactic-disambiguation-modern/).

### Dialects and social media (2011–2019)

From about 2011 attention shifted to dialectal Arabic, driven by user-generated text online. Key steps were the [Arabic Online Commentary Dataset](https://md.research.elamur.ai/free/en/papers/zaidan-2011-arabic-online-commentary-dataset-annotated/), a proposed spelling convention for dialects in [Conventional Orthography for Dialectal Arabic](https://md.research.elamur.ai/free/en/papers/habash-2012-conventional-orthography-dialectal-arabic/), crowd-sourced parallel data in [Machine Translation of Arabic Dialects](https://md.research.elamur.ai/free/en/papers/zbib-2012-machine-translation-arabic-dialects/), and shared tasks on dialect identification and Arabic sentiment.

### Pretrained models and LLMs (2020 onwards)

The arrival of BERT in 2018 was followed by a wave of Arabic and Hebrew pretrained encoders: [AraBERT](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/), [ARBERT & MARBERT](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/), CAMeLBERT, QARiB, [AlephBERT](https://md.research.elamur.ai/free/en/papers/seker-2021-alephbert-hebrew-large-pre-trained/) and [DictaBERT](https://md.research.elamur.ai/free/en/papers/shmidman-2023-dictabert-state-art-bert-suite/). From 2023 the focus moved to generative models trained with Arabic as a first-class language, such as [Jais](https://md.research.elamur.ai/free/en/papers/sengupta-2023-jais-jais-chat-arabic-centric/), [AceGPT](https://md.research.elamur.ai/free/en/papers/huang-2024-acegpt-localizing-large-language-models/), [ALLaM](https://md.research.elamur.ai/free/en/papers/bari-2024-allam-large-language-models-arabic/) and [Fanar](https://md.research.elamur.ai/free/en/papers/abbas-2025-fanar-arabic-centric-multimodal-generative/), and on Hebrew with the DictaLM series.

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 2005–2021; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [2005 — Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop](https://md.research.elamur.ai/free/en/papers/habash-2005-arabic-tokenization-part-speech-tagging/) | 635 |
| [2005 — Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation](https://md.research.elamur.ai/free/en/papers/koehn-2005-edinburgh-system-description-2005-iwslt/) | 545 |
| [2014 — MADAMIRA: A Fast, Comprehensive Tool for Morphological Analysis and Disambiguation of Arabic](https://md.research.elamur.ai/free/en/papers/pasha-2014-madamira-fast-comprehensive-tool-morphological/) | 836 |
| [2014 — Arabic Dialect Identification](https://md.research.elamur.ai/free/en/papers/zaidan-2014-arabic-dialect-identification/) | 437 |
| [2015 — ASTD: Arabic Sentiment Tweets Dataset](https://md.research.elamur.ai/free/en/papers/nabil-2015-astd-arabic-sentiment-tweets-dataset/) | 470 |
| [2016 — Farasa: A Fast and Furious Segmenter for Arabic](https://md.research.elamur.ai/free/en/papers/abdelali-2016-farasa-fast-furious-segmenter-arabic/) | 507 |
| [2017 — SemEval-2017 Task 4: Sentiment Analysis in Twitter](https://md.research.elamur.ai/free/en/papers/rosenthal-2017-semeval-2017-task-4-sentiment/) | 1,161 |
| [2017 — Abusive Language Detection on Arabic Social Media](https://md.research.elamur.ai/free/en/papers/mubarak-2017-abusive-language-detection-arabic-social/) | 414 |
| [2018 — Colorless Green Recurrent Networks Dream Hierarchically](https://md.research.elamur.ai/free/en/papers/gulordava-2018-colorless-green-recurrent-networks-dream/) | 760 |
| [2020 — AraBERT: Transformer-based Model for Arabic Language Understanding](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/) | 1,302 |
| [2020 — KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media](https://md.research.elamur.ai/free/en/papers/safaya-2020-kuisail-semeval-2020-task-12/) | 410 |
| [2021 — ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/) | 478 |

Papers on this topic by year 1,483 dated papers from 1969 to 2026; the busiest year is 2026 with 206, and 74% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 206 |
| 2025 | 174 |
| 2024 | 174 |
| 2023 | 111 |
| 2022 | 111 |
| 2021 | 114 |
| 2020 | 124 |
| 2019 | 52 |
| 2018 | 37 |
| 2017 | 34 |
| 2016 | 46 |
| 2015 | 22 |
| 2014 | 32 |
| 2013 | 15 |
| 2012 | 50 |
| 2011 | 19 |
| 2010 | 27 |
| 2009 | 14 |
| 2008 | 22 |
| 2007 | 8 |
| 2006 | 15 |
| 2005 | 7 |
| 2004 | 17 |
| 2003 | 5 |
| 2002 | 8 |
| 2001 | 2 |
| 2000 | 6 |
| 1999 | 1 |
| 1998 | 8 |
| 1996 | 1 |
| 1995 | 1 |
| 1994 | 3 |
| 1993 | 1 |
| 1991 | 3 |
| 1990 | 1 |
| 1988 | 2 |
| 1982 | 1 |
| 1979 | 2 |
| 1977 | 2 |
| 1976 | 2 |
| 1974 | 2 |
| 1969 | 1 |

## Main problems and tasks

### Morphological analysis and disambiguation

The central task is to assign each word its full analysis: segmentation into clitics and stem, lemma, part of speech, and features such as gender, number, person, aspect, voice, mood, case and state. Because an analyzer typically proposes several readings per word, the problem splits into analysis (listing candidates) and disambiguation (choosing one in context). Dialects multiply the difficulty: [MAGEAD](https://md.research.elamur.ai/free/en/papers/habash-2006-magead-morphological-analyzer-generator-arabic/), [A Morphological Analyzer for Egyptian Arabic](https://md.research.elamur.ai/free/en/papers/habash-2012-morphological-analyzer-egyptian-arabic/) and [Morphological Analysis and Disambiguation for Gulf Arabic](https://md.research.elamur.ai/free/en/papers/khalifa-2020-morphological-analysis-disambiguation-gulf-arabic/) each extend the approach to new varieties. Root extraction is a related task, studied in [Identifying Semitic Roots](https://md.research.elamur.ai/free/en/papers/daya-2008-identifying-semitic-roots-machine-learning/) and [morphologically sensitive clustering of Arabic roots](https://md.research.elamur.ai/free/en/papers/roeck-2000-morphologically-sensitive-clustering-algorithm-identifying/).

### Segmentation and tokenization

Splitting clitics off the stem helps nearly every downstream system, from machine translation to retrieval. [Farasa](https://md.research.elamur.ai/free/en/papers/abdelali-2016-farasa-fast-furious-segmenter-arabic/) made this fast; [A Neural Architecture for Dialectal Arabic Segmentation](https://md.research.elamur.ai/free/en/papers/samih-2017-neural-architecture-dialectal-arabic-segmentation/) carried it to dialects. With neural models, the question became how linguistic segmentation interacts with subword vocabularies, as tested in [Evaluating Various Tokenizers for Arabic Text Classification](https://md.research.elamur.ai/free/en/papers/alyafeai-2021-evaluating-various-tokenizers-arabic-text/) and [Challenging Language-Dependent Segmentation for Arabic](https://md.research.elamur.ai/free/en/papers/sajjad-2017-challenging-language-dependent-segmentation-arabic/).

### Diacritization

Restoring the omitted vowels is needed for text-to-speech, language learning, and some search and translation tasks. Arabic case endings are the hardest part, since they depend on syntax. Approaches run from weighted transducers and maximum-entropy taggers to recurrent networks; see [Arabic Diacritization with Recurrent Neural Networks](https://md.research.elamur.ai/free/en/papers/belinkov-2015-arabic-diacritization-recurrent-neural-networks/) and [Arabic Text Diacritization Using Deep Neural Networks](https://md.research.elamur.ai/free/en/papers/fadel-2019-arabic-text-diacritization-using-deep/). For Hebrew, [Nakdan: Professional Hebrew Diacritizer](https://md.research.elamur.ai/free/en/papers/shmidman-2020-nakdan-professional-hebrew-diacritizer/) addresses vocalization of modern and rabbinic text.

### Parsing

Constituency and dependency parsing of Arabic and Hebrew must cope with segmentation errors, rich agreement and free word order. Work here includes [Better Arabic Parsing](https://md.research.elamur.ai/free/en/papers/green-2010-better-arabic-parsing-baselines-evaluations/), [a study of morphological features in MSA dependency parsing](https://md.research.elamur.ai/free/en/papers/marton-2013-dependency-parsing-modern-standard-arabic/), [Easy-First Dependency Parsing of Modern Hebrew](https://md.research.elamur.ai/free/en/papers/goldberg-2010-easy-first-dependency-parsing-modern/) and joint lattice models such as [a graph-based lattice dependency parser](https://md.research.elamur.ai/free/en/papers/seeker-2015-graph-based-lattice-dependency-parser/).

### Dialect identification

Telling Egyptian from Gulf, or Cairo from Alexandria, from a short sentence is hard because all varieties share a script and much vocabulary. The task moved from five regional classes in [Arabic Dialect Identification](https://md.research.elamur.ai/free/en/papers/zaidan-2014-arabic-dialect-identification/) to city-level distinctions in [Fine-Grained Arabic Dialect Identification](https://md.research.elamur.ai/free/en/papers/salameh-2018-fine-grained-arabic-dialect-identification/).

### Machine translation

Arabic–English was one of the main language pairs of statistical MT. Work here explored how preprocessing affects translation, how to keep several analyses alive during decoding, and how to translate dialects with little parallel data. See [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) for the wider field.

### Social media, sentiment and safety

Sentiment analysis, sarcasm detection, offensive language and hate speech, fake news and stance detection make up a large and fast-growing share of recent work, almost all of it on Arabic Twitter. Representative datasets are [ASTD](https://md.research.elamur.ai/free/en/papers/nabil-2015-astd-arabic-sentiment-tweets-dataset/), [ArSarcasm](https://md.research.elamur.ai/free/en/papers/farha-2020-arabic-sentiment-analysis-sarcasm-detection/), [AraStance](https://md.research.elamur.ai/free/en/papers/alhindi-2021-arastance-multi-country-multi-domain/) and [ArMIS](https://md.research.elamur.ai/free/en/papers/almanea-2022-armis-arabic-misogyny-sexism-corpus/).

### Speech

Arabic speech recognition faces the same diglossia as text, plus code-switching between dialect, MSA, English and French. The MGB challenges on broadcast media and studies such as [a comparison of end-to-end, modular and human Arabic speech recognition](https://md.research.elamur.ai/free/en/papers/hussein-2021-arabic-speech-recognition-end-end/) frame much of this work; Amharic appears in [an Amharic speech emotion dataset](https://md.research.elamur.ai/free/en/papers/retta-2022-new-amharic-speech-emotion-dataset/).

### Ancient and historical languages

Computational methods are increasingly applied to ancient Semitic texts: restoring broken cuneiform tablets in [restoring fragmentary Babylonian texts with recurrent networks](https://md.research.elamur.ai/free/en/papers/fetaya-2020-restoration-fragmentary-babylonian-texts-using/) and [Filling the Gaps in Ancient Akkadian Texts](https://md.research.elamur.ai/free/en/papers/lazar-2021-filling-gaps-ancient-akkadian-texts/), modelling Babylonian morphology in [BabyFST](https://md.research.elamur.ai/free/en/papers/sahala-2020-babyfst-finite-state-based-computational/), and automatic decipherment in [neural decipherment of Ugaritic and Linear B](https://md.research.elamur.ai/free/en/papers/luo-2019-neural-decipherment-minimum-cost-flow/).

## Methods and schools

### Rule-based and finite-state

The earliest school treated morphology as a formal system to be written down. Multi-tape transducers, two-level rules and lexicon-plus-rules analyzers produce every legal reading of a word and never produce an illegal one. Their weakness is coverage and the cost of writing rules; their strength is that their output is always a well-formed analysis. Later systems such as [ElixirFM](https://md.research.elamur.ai/free/en/papers/smrz-2007-elixirfm-implementation-functional-arabic-morphology/), the [open-source finite-state transducer for MSA](https://md.research.elamur.ai/free/en/papers/attia-2011-open-source-finite-state-morphological/) and [CALIMA-Star, an analyzer with copious features](https://md.research.elamur.ai/free/en/papers/taji-2018-arabic-morphological-analyzer-generator-copious/) keep that tradition alive.

### Analyze-and-rank

The most influential school in Arabic NLP combined a rule-based analyzer with statistical classifiers. The analyzer lists candidates; separate classifiers predict individual features; a scoring step picks the candidate that agrees best with the predictions. This design, introduced by Habash and Rambow in 2005, runs through MADA, MADAMIRA, the neural disambiguator of [Don't Throw Those Morphological Analyzers Away](https://md.research.elamur.ai/free/en/papers/zalmout-2017-don-t-throw-those-morphological/) and [CAMeL Tools](https://md.research.elamur.ai/free/en/papers/obeid-2020-camel-tools-open-source-python/). Its appeal is that the output always comes from a linguistically valid inventory, while the choice among readings is learned from data.

### Analyzer-free and lightweight

A competing line, associated with Qatar-based researchers, argued that a full analyzer is not always worth its cost. Farasa ranks candidate segmentations with a linear model and a few lexicons; Darwish's early root analyser was learned from word-root pairs in under a day. These tools trade some depth for speed and simplicity, and in tasks like retrieval and translation the difference in quality is small.

### Joint and lattice models

Hebrew researchers pushed hardest on the idea that segmentation and syntax should be decided together rather than in a pipeline. Lattices of possible segmentations are passed to the parser or to the decoder, which picks the path and the structure at once. The same idea appears in machine translation in [Generalizing Word Lattice Translation](https://md.research.elamur.ai/free/en/papers/dyer-2008-generalizing-word-lattice-translation/), and in later transition-based work such as [Joint Transition-Based Models for Morpho-Syntactic Parsing](https://md.research.elamur.ai/free/en/papers/more-2019-joint-transition-based-models-morpho/).

### Neural and pretrained

Since about 2015, recurrent networks and then transformers have taken over most tasks. Two debates run through this work. The first is what to pretrain on: standard Arabic, dialects, tweets or classical text, and in what mixture; a [controlled comparison of Arabic pretrained models](https://md.research.elamur.ai/free/en/papers/inoue-2021-interplay-variant-size-task-type/) found that the match between pretraining variety and task matters more than raw size. The second is how to tokenize a language whose words pack several morphemes; Hebrew work such as [Large Pre-Trained Models with Extra-Large Vocabularies](https://md.research.elamur.ai/free/en/papers/gueta-2022-large-pre-trained-models-extra/) argues for much larger vocabularies, while Arabic work has experimented with morphological pre-segmentation before subword splitting.

### Building for culture, not only language

With large language models, a new concern appeared: a model can write fluent Arabic while defaulting to Western cultural assumptions. [Having Beer after Prayer?](https://md.research.elamur.ai/free/en/papers/naous-2023-having-beer-after-prayer-measuring/) made this measurable, and several Arabic-centric LLMs now describe cultural alignment as a design goal alongside language quality.

## Datasets, resources and benchmarks

| Resource | Kind | Where to read about it |
| --- | --- | --- |
| Penn Arabic Treebank | Morphologically and syntactically annotated MSA news | [Developing an Arabic Treebank](https://md.research.elamur.ai/free/en/papers/maamouri-2004-developing-arabic-treebank-methods-guidelines/) |
| Columbia Arabic Treebank (CATiB) | Dependency treebank with a small tag set | [CATiB](https://md.research.elamur.ai/free/en/papers/habash-2009-catib-columbia-arabic-treebank/) |
| Quranic Arabic Corpus | Morphology and syntax of the Qur'an | [Morphological Annotation of Quranic Arabic](https://md.research.elamur.ai/free/en/papers/dukes-2010-morphological-annotation-quranic-arabic/) |
| Universal Dependencies treebanks | Cross-linguistically consistent trees for Arabic and Hebrew | [Universal Dependencies for Arabic](https://md.research.elamur.ai/free/en/papers/taji-2017-universal-dependencies-arabic/), [The Hebrew Universal Dependency Treebank](https://md.research.elamur.ai/free/en/papers/sade-2018-hebrew-universal-dependency-treebank-past/) |
| Arabic Online Commentary | Dialect-labelled newspaper comments | [AOC](https://md.research.elamur.ai/free/en/papers/zaidan-2011-arabic-online-commentary-dataset-annotated/) |
| MADAR | Parallel city-level dialect corpus and dialect-labelled tweets | [The MADAR Shared Task](https://md.research.elamur.ai/free/en/papers/bouamor-2019-madar-shared-task-arabic-fine/) |
| Multidialectal parallel corpus | Aligned sentences in several dialects, MSA and English | [A Multidialectal Parallel Corpus of Arabic](https://md.research.elamur.ai/free/en/papers/bouamor-2014-multidialectal-parallel-corpus-arabic/) |
| QALB | Corrected Arabic text for error correction | [Large Scale Arabic Error Annotation](https://md.research.elamur.ai/free/en/papers/zaghouani-2014-large-scale-arabic-error-annotation/) |
| MGB-2 and MGB-3 | Arabic broadcast speech | [The MGB-2 Challenge](https://md.research.elamur.ai/free/en/papers/ali-2016-mgb-2-challenge-arabic-multi/), [Arabic MGB-3](https://md.research.elamur.ai/free/en/papers/ali-2017-speech-recognition-challenge-wild-arabic/) |
| ArabicMMLU | Native-Arabic exam questions for LLMs | [ArabicMMLU](https://md.research.elamur.ai/free/en/papers/koto-2024-arabicmmlu-assessing-massive-multitask-language/) |
| Wojood | Nested named entities | [Wojood](https://md.research.elamur.ai/free/en/papers/jarrar-2022-wojood-nested-arabic-named-entity/) |

Beyond these, the field relies on shared tasks: the SemEval Twitter sentiment task with an Arabic track, the VarDial and MADAR dialect tasks, the WANLP sarcasm task, OffensEval, and the NADI series. Catalogues of resources such as [Masader](https://md.research.elamur.ai/free/en/papers/alyafeai-2022-masader-metadata-sourcing-arabic-text/) and evaluation suites such as [LAraBench](https://md.research.elamur.ai/free/en/papers/abdelali-2024-larabench-benchmarking-arabic-ai-large/) help researchers find and compare them. For Hebrew, the Hebrew Treebank and its UD conversion, and the evaluation suites introduced with AlephBERT, play the corresponding role; [What's Wrong with Hebrew NLP?](https://md.research.elamur.ai/free/en/papers/tsarfaty-2019-what-s-wrong-hebrew-nlp/) set out what was missing.

## Landmark papers in this library

### AraBERT (2020)

[AraBERT: Transformer-based Model for Arabic Language Understanding](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/) trained a BERT-base model on Arabic text alone, segmenting words with Farasa before building the subword vocabulary. It outperformed multilingual BERT on sentiment, named entities and question answering, and became the default starting point for Arabic classification work. It is the most cited work on this topic here, and nearly every later Arabic encoder compares against it.

### MADAMIRA (2014)

[MADAMIRA](https://md.research.elamur.ai/free/en/papers/pasha-2014-madamira-fast-comprehensive-tool-morphological/) merged the morphological disambiguator MADA and the shallow processor AMIRA into a single Java system with a server mode. It covered both MSA and Egyptian, offered many tokenization schemes from one chosen analysis, and ran more than an order of magnitude faster than its predecessors with almost no loss in accuracy. For years it was the standard preprocessing step for Arabic MT, parsing and information extraction.

### One fell swoop (2005)

[Habash and Rambow's "One Fell Swoop" paper](https://md.research.elamur.ai/free/en/papers/habash-2005-arabic-tokenization-part-speech-tagging/) introduced the analyze-and-rank method: SVM classifiers predict individual features and the best-agreeing analysis from the Buckwalter-based analyzer wins. It showed that a morphological analyzer is more valuable for richly inflected languages than extra annotated data, and it became the foundation of MADA.

### Farasa (2016)

[Farasa](https://md.research.elamur.ai/free/en/papers/abdelali-2016-farasa-fast-furious-segmenter-arabic/) ranked candidate segmentations with a linear model and a few lexicons, matching MADAMIRA's segmentation accuracy while running dozens of times faster. Its speed made it the segmenter of choice for large-scale work, including the pretraining corpus of AraBERT.

### ARBERT and MARBERT (2021)

[ARBERT & MARBERT](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/) trained two models, one on diverse MSA text and one on a very large collection of dialect-rich tweets, and evaluated them on a broad benchmark of Arabic understanding tasks. MARBERT became a strong default for social-media and dialect work.

### Arabic Dialect Identification (2014)

[Arabic Dialect Identification](https://md.research.elamur.ai/free/en/papers/zaidan-2014-arabic-dialect-identification/) built on the Arabic Online Commentary data, with crowd annotators marking, sentence by sentence, the degree of dialect and its region. It established language-model classifiers as a baseline and documented annotators' tendency to attribute sentences to their own dialect.

### Abusive Language Detection on Arabic Social Media (2017)

[Abusive Language Detection on Arabic Social Media](https://md.research.elamur.ai/free/en/papers/mubarak-2017-abusive-language-detection-arabic-social/) mined obscene vocabulary by contrasting the word statistics of users who do and do not use known insult terms. It opened a line of Arabic offensive-language research that continued with [Arabic Offensive Language on Twitter](https://md.research.elamur.ai/free/en/papers/mubarak-2021-arabic-offensive-language-twitter-analysis/) and [emoji-anchored hate-speech collection](https://md.research.elamur.ai/free/en/papers/mubarak-2022-emojis-anchors-detect-arabic-offensive/), each finding a cheaper way to collect a high share of offensive tweets.

### CAMeLBERT: variant, size and task (2021)

[This controlled study of Arabic pretrained models](https://md.research.elamur.ai/free/en/papers/inoue-2021-interplay-variant-size-task-type/) released the CAMeLBERT family, pretrained under controlled conditions on MSA, dialectal, classical and mixed data. Its finding that the pretraining variety should match the task data more than it should be large changed how Arabic models were chosen.

### Arabic Preprocessing Schemes for Statistical Machine Translation (2006)

[Arabic Preprocessing Schemes for Statistical Machine Translation](https://md.research.elamur.ai/free/en/papers/habash-2006-arabic-preprocessing-schemes-statistical-machine/) separated the question of how to split Arabic words from the question of which tool does the splitting, and measured both against training-set size. Its result, that the best scheme depends on how much data is available, guided Arabic MT preprocessing for years.

### CAMeL Tools (2020)

[CAMeL Tools](https://md.research.elamur.ai/free/en/papers/obeid-2020-camel-tools-open-source-python/) gathered morphological analysis, disambiguation, dialect identification, named entity recognition and sentiment into one open-source Python package from NYU Abu Dhabi. It replaced a patchwork of tools written in different languages with different licences.

### Conventional Orthography for Dialectal Arabic (2012)

[Conventional Orthography for Dialectal Arabic](https://md.research.elamur.ai/free/en/papers/habash-2012-conventional-orthography-dialectal-arabic/) proposed CODA, a spelling convention for dialects designed for computational use, starting with Egyptian. It was generalised across dialects as [CODA*](https://md.research.elamur.ai/free/en/papers/habash-2018-unified-guidelines-resources-arabic-dialect/) and underlies many dialect resources.

### Having Beer after Prayer? (2023)

[Having Beer after Prayer?](https://md.research.elamur.ai/free/en/papers/naous-2023-having-beer-after-prayer-measuring/) built a set of Arab and Western entities and natural prompts to test whether language models trained with Arabic text favour Western content. They do, across generation, recognition and sentiment tasks, which made cultural fit a measurable property of Arabic models.

### Jais (2023)

[Jais and Jais-chat](https://md.research.elamur.ai/free/en/papers/sengupta-2023-jais-jais-chat-arabic-centric/) released an open 13-billion-parameter decoder trained on Arabic and English with an instruction-tuned chat version. It was among the first openly available generative models built with Arabic as a primary language, and was followed by AceGPT, ALLaM and Fanar.

Citations of the papers on this topic, by when they were published Papers published in 2020–2024 gather 37% of the 31,207 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 380 papers, 726 citations |
| 2020–2024 | 634 papers, 11,575 citations |
| 2015–2019 | 191 papers, 7,944 citations |
| 2010–2014 | 143 papers, 5,172 citations |
| 2005–2009 | 66 papers, 4,230 citations |
| 2000–2004 | 38 papers, 1,273 citations |
| 1995–1999 | 11 papers, 215 citations |
| 1990–1994 | 8 papers, 65 citations |
| 1985–1989 | 2 papers, 7 citations |
| 1980–1984 | 1 papers, 0 citations |
| 1975–1979 | 6 papers, 0 citations |
| 1970–1974 | 2 papers, 0 citations |
| 1965–1969 | 1 papers, 0 citations |

## People

[Nizar Habash](https://md.research.elamur.ai/free/en/authors/nizar-habash/), now at New York University Abu Dhabi, is the most prolific author on this topic. His work spans the analyze-and-rank method, MADA and MADAMIRA, dialect morphology (MAGEAD, CALIMA), the CODA orthography, the CATiB treebank, dialect identification with MADAR, and CAMeL Tools; few areas of Arabic NLP lack a contribution from him.

[Owen Rambow](https://md.research.elamur.ai/free/en/authors/owen-rambow/), now at Stony Brook University, co-authored many of Habash's foundational papers, including the 2005 disambiguation method, MAGEAD, the work on [Parsing Arabic Dialects](https://md.research.elamur.ai/free/en/papers/chiang-2006-parsing-arabic-dialects/) and the study of morphological features in dependency parsing.

[Kareem Darwish](https://md.research.elamur.ai/free/en/authors/kareem-darwish/) of Hamad bin Khalifa University in Qatar began with a quickly built root analyser for retrieval and went on to Farasa, diacritization with [Arabic Diacritization: Stats, Rules, and Hacks](https://md.research.elamur.ai/free/en/papers/darwish-2017-arabic-diacritization-stats-rules-hacks/), and offensive-language detection.

[Hamdy Mubarak](https://md.research.elamur.ai/free/en/authors/hamdy-mubarak/), also at Hamad bin Khalifa University, is behind much of the Arabic social-media work here: abusive and offensive language datasets, the QARiB tweet models, and the MGB-2 speech challenge.

[Ahmed Abdelali](https://md.research.elamur.ai/free/en/authors/ahmed-abdelali/), now at Qatar Science and Technology Park, led Farasa and the study [Pre-Training BERT on Arabic Tweets: Practical Considerations](https://md.research.elamur.ai/free/en/papers/abdelali-2021-pre-training-bert-arabic-tweets/), and co-authored LAraBench.

[Reut Tsarfaty](https://md.research.elamur.ai/free/en/authors/reut-tsarfaty/), now at Google and Bar-Ilan University, is the leading voice in Hebrew parsing. Her papers here range from joint morphological and syntactic disambiguation of Modern Hebrew in 2006 to AlephBERT and an argument for morphology-aware evaluation of Hebrew models.

[Michael Elhadad](https://md.research.elamur.ai/free/en/authors/michael-elhadad/) of Ben-Gurion University of the Negev and [Yoav Goldberg](https://md.research.elamur.ai/free/en/authors/yoav-goldberg/), now at Bar-Ilan University and the Allen Institute for AI, built much of the Hebrew pipeline: unsupervised morpheme-level tagging, easy-first dependency parsing, and [joint Hebrew segmentation and parsing with a lattice parser](https://md.research.elamur.ai/free/en/papers/goldberg-2011-joint-hebrew-segmentation-parsing-pcfgla/).

[Shuly Wintner](https://md.research.elamur.ai/free/en/authors/shuly-wintner/) of the University of Haifa contributed a finite-state grammar of Hebrew, work on morphological disambiguation, root identification with linguistic constraints, and an early Hebrew-to-English transfer translation system.

[Avi Shmidman](https://md.research.elamur.ai/free/en/authors/avi-shmidman/) of Bar-Ilan University, with [Shaltiel Shmidman](https://md.research.elamur.ai/free/en/authors/shaltiel-shmidman/) and [Moshe Koppel](https://md.research.elamur.ai/free/en/authors/moshe-koppel/), produced the Dicta line of Hebrew tools: the Nakdan diacritizer, the DictaBERT encoders and the DictaLM generative models.

[Houda Bouamor](https://md.research.elamur.ai/free/en/authors/houda-bouamor/) of Carnegie Mellon University Qatar created the multidialectal parallel corpus and led the MADAR work on city-level dialect identification.

[Wajdi Zaghouani](https://md.research.elamur.ai/free/en/authors/wajdi-zaghouani/) of Northwestern University in Qatar designed annotation guidelines for large resources, among them the QALB error-correction corpus, and co-organised the Arabic sarcasm shared task.

[Bashar Alhafni](https://md.research.elamur.ai/free/en/authors/bashar-alhafni/), now affiliated with Cambridge and with MBZUAI in Abu Dhabi, and [Salam Khalifa](https://md.research.elamur.ai/free/en/authors/salam-khalifa/), now at New York University Abu Dhabi, contributed to CAMeL Tools and CAMeLBERT; Alhafni's own work covers grammatical error correction and text simplification, and Khalifa's covers Gulf Arabic morphology and dialect orthography.

[Ahmed Ali](https://md.research.elamur.ai/free/en/authors/ahmed-ali/) led the Arabic multi-genre broadcast speech challenges, co-organised the VarDial task that included Arabic dialect identification, and has worked on code-switching and end-to-end speech recognition.

[Walid Magdy](https://md.research.elamur.ai/free/en/authors/walid-magdy/) of the University of Edinburgh and [Preslav Nakov](https://md.research.elamur.ai/free/en/authors/preslav-nakov/), now at MBZUAI in Abu Dhabi, appear mainly through shared tasks and datasets: Nakov through SemEval sentiment and ArabicMMLU, Magdy through ArSarcasm and iSarcasmEval.

[Mustafa Jarrar](https://md.research.elamur.ai/free/en/authors/mustafa-jarrar/) built the Arabic Ontology and the nested named-entity corpus Wojood, while [Zaid Alyafeai](https://md.research.elamur.ai/free/en/authors/zaid-alyafeai/) works on resource catalogues, tokenizer evaluation and Arabic poetry.

Authors with the most papers on this topic Nizar Habash leads with 83 papers; the 12 authors shown wrote 293 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Nizar Habash](https://md.research.elamur.ai/free/en/authors/nizar-habash/) | 83 papers · 7,119 citations |
| [Hamdy Mubarak](https://md.research.elamur.ai/free/en/authors/hamdy-mubarak/) | 28 papers · 2,573 citations |
| [Reut Tsarfaty](https://md.research.elamur.ai/free/en/authors/reut-tsarfaty/) | 23 papers · 671 citations |
| [Avi Shmidman](https://md.research.elamur.ai/free/en/authors/avi-shmidman/) | 22 papers · 144 citations |
| [Wajdi Zaghouani](https://md.research.elamur.ai/free/en/authors/wajdi-zaghouani/) | 21 papers · 763 citations |
| [Kareem Darwish](https://md.research.elamur.ai/free/en/authors/kareem-darwish/) | 19 papers · 1,990 citations |
| [Owen Rambow](https://md.research.elamur.ai/free/en/authors/owen-rambow/) | 18 papers · 3,173 citations |
| [Ahmed Abdelali](https://md.research.elamur.ai/free/en/authors/ahmed-abdelali/) | 17 papers · 1,403 citations |
| [Salam Khalifa](https://md.research.elamur.ai/free/en/authors/salam-khalifa/) | 16 papers · 651 citations |
| [Shammur Chowdhury](https://md.research.elamur.ai/free/en/authors/shammur-absar-chowdhury/) | 16 papers · 627 citations |
| [Houda Bouamor](https://md.research.elamur.ai/free/en/authors/houda-bouamor/) | 15 papers · 1,498 citations |
| [Muhammad Abdul-Mageed](https://md.research.elamur.ai/free/en/authors/muhammad-abdul-mageed/) | 15 papers · 1,381 citations |

## Institutions and places

Computational work on this topic is concentrated in a few clusters. The largest is the CAMeL Lab at New York University Abu Dhabi and its New York parent; its alumni now work across the Gulf, Europe and North America. The second is in Qatar: Hamad bin Khalifa University, where the Qatar Computing Research Institute has produced Farasa, QARiB, the MGB challenges and the Fanar platform, together with Carnegie Mellon University Qatar and Northwestern University in Qatar. Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi is a newer centre for Arabic LLMs and benchmarks.

Israeli universities form the Hebrew cluster: Bar-Ilan University (the Dicta group and Tsarfaty's lab), the Hebrew University of Jerusalem, Tel Aviv University, Ben-Gurion University of the Negev and the University of Haifa. In Saudi Arabia, King Fahd University of Petroleum and Minerals, King Abdullah University of Science and Technology and Prince Sultan University have become active in datasets, embeddings and LLMs. In the United States, New York University, Carnegie Mellon, Stony Brook and Google appear through the careers of leading authors, and in Europe the universities of Edinburgh, Cambridge and Hamburg.

By country, authors working in the United States lead, followed by Israel, the United Arab Emirates, the United Kingdom, Saudi Arabia and Qatar. The venues reflect the two traditions: the Arabic NLP workshop and conference series, LREC and SemEval on the computational side, and journals such as Glossa, Aethiopica and Folia Orientalia on the linguistic and philological side.

Where the authors on this topic work now New York University Abu Dhabi is the current home of authors of 93 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [New York University Abu Dhabi](https://md.research.elamur.ai/free/en/universities/new-york-university-abu-dhabi/) | 93 papers |
| [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/) | 91 papers |
| [Hamad bin Khalifa University](https://md.research.elamur.ai/free/en/universities/hamad-bin-khalifa-university/) | 71 papers |
| [Mohamed bin Zayed University of Artificial Intelligence](https://md.research.elamur.ai/free/en/universities/mohamed-bin-zayed-university-of-artificial-intelligence/) | 56 papers |
| [Bar-Ilan University](https://md.research.elamur.ai/free/en/universities/bar-ilan-university/) | 55 papers |
| [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) | 47 papers |
| [Google](https://md.research.elamur.ai/free/en/companies/google/) | 32 papers |
| [Stony Brook University](https://md.research.elamur.ai/free/en/universities/stony-brook-university/) | 31 papers |
| [Hebrew University of Jerusalem](https://md.research.elamur.ai/free/en/universities/hebrew-university-of-jerusalem/) | 26 papers |
| [University of Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) | 26 papers |

Countries of the authors on this topic United States comes first: authors working there now wrote 308 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| [United States](https://md.research.elamur.ai/free/en/countries/united-states/) | 308 papers |
| [Israel](https://md.research.elamur.ai/free/en/countries/israel/) | 144 papers |
| [United Arab Emirates](https://md.research.elamur.ai/free/en/countries/united-arab-emirates/) | 135 papers |
| [Qatar](https://md.research.elamur.ai/free/en/countries/qatar/) | 105 papers |
| [Saudi Arabia](https://md.research.elamur.ai/free/en/countries/saudi-arabia/) | 100 papers |
| [United Kingdom](https://md.research.elamur.ai/free/en/countries/united-kingdom/) | 99 papers |
| [Germany](https://md.research.elamur.ai/free/en/countries/germany/) | 74 papers |
| [France](https://md.research.elamur.ai/free/en/countries/france/) | 60 papers |
| [Egypt](https://md.research.elamur.ai/free/en/countries/egypt/) | 50 papers |
| [Canada](https://md.research.elamur.ai/free/en/countries/canada/) | 41 papers |

Where the papers on this topic were published Language Resources and Evaluation Conference (LREC) carries the most, 160 of 1485.

**The numbers**

| Venue | Papers |
| --- | --- |
| [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/) | 160 papers |
| [Folia Orientalia](https://md.research.elamur.ai/free/en/venues/folia-orientalia/) | 44 papers |
| [Arabic Natural Language Processing Workshop (WANLP)](https://md.research.elamur.ai/free/en/venues/wanlp/) | 42 papers |
| [Glossa a journal of general linguistics](https://md.research.elamur.ai/free/en/venues/glossa-a-journal-of-general-linguistics/) | 40 papers |
| [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/) | 33 papers |
| [Arabic Natural Language Processing Conference (ArabicNLP)](https://md.research.elamur.ai/free/en/venues/arabicnlp/) | 26 papers |
| [Aethiopica](https://md.research.elamur.ai/free/en/venues/aethiopica/) | 25 papers |
| [Findings of the Association for Computational Linguistics](https://md.research.elamur.ai/free/en/venues/findings/) | 25 papers |
| [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/) | 21 papers |
| [Journal of Hebrew Scriptures](https://md.research.elamur.ai/free/en/venues/journal-of-hebrew-scriptures/) | 21 papers |

## Open questions

- **Dialects remain under-served.** Despite progress, most dialects other than Egyptian and Levantine have little annotated data, and text written in Latin script or mixed with French and English is still hard to handle; see [transfer of multilingual models to North African Arabizi](https://md.research.elamur.ai/free/en/papers/muller-2020-multilingual-language-models-transfer-unseen/).
- **The rest of the family.** Amharic, Tigrinya, Maltese and the Neo-Aramaic languages receive a small fraction of the attention given to Arabic and Hebrew, and ancient languages depend on small, specialised communities.
- **Tokenization for root-and-pattern languages.** Whether subword vocabularies should respect morpheme boundaries, and how large they should be, is not settled; results differ between Arabic and Hebrew and between understanding and generation.
- **Do analyzers still matter?** Neural models perform well without explicit morphology on many tasks, yet morphological analyzers still help in low-data settings and dialects. Where the boundary lies is an open empirical question.
- **Evaluation of LLMs.** Translated benchmarks miss local knowledge and introduce errors; native benchmarks such as ArabicMMLU are a start, but evaluation of generation, dialect fluency and cultural fit is still immature.
- **Annotation subjectivity.** Sentiment, sarcasm and offensiveness labels vary with annotators' dialect and background, and several datasets here now keep disagreements rather than averaging them away.

## Neighbouring topics

The closest neighbour is [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/), which shares about a third of the papers here and covers the ancient Northwest Semitic texts in more depth. [Morphology](https://md.research.elamur.ai/free/en/topics/morphology/) is the next: Semitic root-and-pattern systems are a standard test for any theory or tool of word structure, and [Tokenization & segmentation](https://md.research.elamur.ai/free/en/topics/tokenization/) and [Part-of-speech tagging](https://md.research.elamur.ai/free/en/topics/part-of-speech-tagging/) treat the practical side of the same problem.

Much of the recent Arabic work belongs equally to [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) and [Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/), through social-media sentiment, sarcasm and offensive language. The pretrained models connect the topic to [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) and [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/). Dialects and code-switching tie it to [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/), [Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/) and [Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/). Syntax work links to [Treebanks](https://md.research.elamur.ai/free/en/topics/treebanks/), [Dependency parsing](https://md.research.elamur.ai/free/en/topics/dependency-parsing/), [Universal Dependencies](https://md.research.elamur.ai/free/en/topics/universal-dependencies/) and [Syntax & grammar formalisms](https://md.research.elamur.ai/free/en/topics/syntax-grammar/), and translation work to [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/). Question answering, named entities and multiword expressions appear through [Question answering](https://md.research.elamur.ai/free/en/topics/question-answering/), [Named entity recognition](https://md.research.elamur.ai/free/en/topics/named-entity-recognition/) and [Idioms & multiword expressions](https://md.research.elamur.ai/free/en/topics/multiword-expressions/). The ancient languages connect to [Akkadian and cuneiform NLP](https://md.research.elamur.ai/free/en/topics/akkadian-cuneiform-nlp/) and [Historical linguistics & language change](https://md.research.elamur.ai/free/en/topics/language-change/).

Almost all papers on this topic sit on the [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) shelf; many also sit on [Natural language processing](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) and [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/), and a few on [History & humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/history/) and [Law](https://md.research.elamur.ai/free/en/shelves/social-sciences/law/).

Neighbouring topics: papers shared Hebrew & Aramaic shares the most: 392 of the 1,485 papers here (26%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) | 392 shared · 26% |
| [Biblical Hebrew](https://md.research.elamur.ai/free/en/topics/biblical-hebrew/) | 190 shared · 13% |
| [Morphology](https://md.research.elamur.ai/free/en/topics/morphology/) | 157 shared · 11% |
| [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) | 145 shared · 10% |
| [Modern Hebrew](https://md.research.elamur.ai/free/en/topics/modern-hebrew/) | 138 shared · 9% |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 99 shared · 7% |
| [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) | 78 shared · 5% |
| [Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) | 77 shared · 5% |
| [Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) | 59 shared · 4% |
| [Aramaic](https://md.research.elamur.ai/free/en/topics/aramaic/) | 57 shared · 4% |

The shelves this topic stands on Language and Linguistics: 1338 of the 1,485 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) | 1,338 papers |
| [Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) | 1,049 papers |
| [Religious studies](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/religious-studies/) | 151 papers |
| [Experimental and Cognitive Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/experimental-and-cognitive-psychology/) | 111 papers |
| [Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) | 90 papers |
| [Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) | 84 papers |
| [Literature and Literary Theory](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/literature-and-literary-theory/) | 78 papers |
| [Education](https://md.research.elamur.ai/free/en/shelves/social-sciences/education/) | 76 papers |
| [Developmental and Educational Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/developmental-and-educational-psychology/) | 64 papers |
| [Sociology and Political Science](https://md.research.elamur.ai/free/en/shelves/social-sciences/sociology-and-political-science/) | 60 papers |
| [Information Systems](https://md.research.elamur.ai/free/en/shelves/computer-science/information-systems/) | 57 papers |
| [Archeology](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/archeology/) | 43 papers |

## Key works in this library

- [Multi-Tape Two-Level Morphology](https://md.research.elamur.ai/free/en/papers/kiraz-1994-multi-tape-two-level-morphology/) — root, pattern and vowels on separate tapes of one automaton.
- [Sebawai](https://md.research.elamur.ai/free/en/papers/darwish-2002-building-shallow-arabic-morphological-analyser/) — Darwish's root analyser learned from word-root pairs, tested in retrieval.
- [Developing an Arabic Treebank](https://md.research.elamur.ai/free/en/papers/maamouri-2004-developing-arabic-treebank-methods-guidelines/) — how the Penn Arabic Treebank was built.
- [One Fell Swoop](https://md.research.elamur.ai/free/en/papers/habash-2005-arabic-tokenization-part-speech-tagging/) — Habash and Rambow's analyze-and-rank method.
- [Arabic Diacritization through Full Morphological Tagging](https://md.research.elamur.ai/free/en/papers/habash-2007-arabic-diacritization-through-full-morphological/) — choosing whole diacritized analyses instead of single marks.
- [CATiB: The Columbia Arabic Treebank](https://md.research.elamur.ai/free/en/papers/habash-2009-catib-columbia-arabic-treebank/) — a fast-to-annotate dependency treebank.
- [Better Arabic Parsing](https://md.research.elamur.ai/free/en/papers/green-2010-better-arabic-parsing-baselines-evaluations/) — why Arabic constituency parsing lagged and how to narrow the gap.
- [The Arabic Online Commentary Dataset](https://md.research.elamur.ai/free/en/papers/zaidan-2011-arabic-online-commentary-dataset-annotated/) — reader comments as a source of written dialect.
- [A Morphological Analyzer for Egyptian Arabic](https://md.research.elamur.ai/free/en/papers/habash-2012-morphological-analyzer-egyptian-arabic/) — CALIMA, the first broad analyzer for a dialect.
- [MADAMIRA](https://md.research.elamur.ai/free/en/papers/pasha-2014-madamira-fast-comprehensive-tool-morphological/) — the standard Arabic preprocessing toolkit of the 2010s.
- [A Multidialectal Parallel Corpus of Arabic](https://md.research.elamur.ai/free/en/papers/bouamor-2014-multidialectal-parallel-corpus-arabic/) — the same sentences across dialects, MSA and English.
- [Farasa](https://md.research.elamur.ai/free/en/papers/abdelali-2016-farasa-fast-furious-segmenter-arabic/) — fast, accurate Arabic segmentation.
- [Fine-Grained Arabic Dialect Identification](https://md.research.elamur.ai/free/en/papers/salameh-2018-fine-grained-arabic-dialect-identification/) — dialect identification at the level of cities.
- [AraBERT](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/) — the first widely used Arabic BERT.
- [AraT5](https://md.research.elamur.ai/free/en/papers/nagoudi-2022-arat5-text-text-transformers-arabic/) — Arabic sequence-to-sequence models and a generation benchmark.
- [AlephBERT (ACL 2022)](https://md.research.elamur.ai/free/en/papers/seker-2022-alephbert-language-model-pre-training/) — a Hebrew encoder with morphology-level evaluation.
- [AceGPT, Localizing Large Language Models in Arabic](https://md.research.elamur.ai/free/en/papers/huang-2024-acegpt-localizing-large-language-models/) — adapting an open LLM to Arabic language and values.
- [ArabicMMLU](https://md.research.elamur.ai/free/en/papers/koto-2024-arabicmmlu-assessing-massive-multitask-language/) — a native Arabic knowledge benchmark built from school and professional exams.

## Papers (1,485)

**1** [2](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/2/) … [10](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/10/) … [16](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/16/) [next →](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/2/)

1. [**AraBERT: Transformer-based Model for Arabic Language Understanding arXiv:2003.00104**Wissam Antoun, Fady Baly, Hazem Hajj*2020 · 1,302 citations*](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/)
2. **[SemEval-2017 Task 4: Sentiment Analysis in Twitter](https://md.research.elamur.ai/free/en/papers/rosenthal-2017-semeval-2017-task-4-sentiment/) [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/)**Sara Rosenthal, Noura Farra, Preslav Nakov*2017 · 1,161 citations*
3. [**MADAMIRA: A Fast, Comprehensive Tool for Morphological Analysis and Disambiguation of Arabic**Arfath Pasha, Mohamed Al-Badrashiny, Mona Diab, et al.*2014 · 836 citations*](https://md.research.elamur.ai/free/en/papers/pasha-2014-madamira-fast-comprehensive-tool-morphological/)
4. **[Colorless Green Recurrent Networks Dream Hierarchically](https://md.research.elamur.ai/free/en/papers/gulordava-2018-colorless-green-recurrent-networks-dream/) [Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)](https://md.research.elamur.ai/free/en/venues/naacl/)**Kristina Gulordava, Piotr Bojanowski, Edouard Grave, et al.*2018 · 760 citations*
5. [**Arabic Tokenization, Part-of-Speech Tagging and Morphological Disambiguation in One Fell Swoop**Nizar Habash, Owen Rambow*2005 · 635 citations*](https://md.research.elamur.ai/free/en/papers/habash-2005-arabic-tokenization-part-speech-tagging/)
6. **[Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation](https://md.research.elamur.ai/free/en/papers/koehn-2005-edinburgh-system-description-2005-iwslt/) [International Conference on Spoken Language Translation (IWSLT)](https://md.research.elamur.ai/free/en/venues/iwslt/)**Philipp Koehn, Amittai Axelrod, Alexandra Birch Mayne, et al.*2005 · 545 citations*
7. [**Farasa: A Fast and Furious Segmenter for Arabic**Ahmed Abdelali, Kareem Darwish, Nadir Durrani, et al.*2016 · 507 citations*](https://md.research.elamur.ai/free/en/papers/abdelali-2016-farasa-fast-furious-segmenter-arabic/)
8. [**ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic arXiv:2101.01785 · ACL-2021 camera ready version**Muhammad Abdul-Mageed, AbdelRahim Elmadany, El Moatez Billah Nagoudi*2021 · 478 citations*](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/)
9. **[ASTD: Arabic Sentiment Tweets Dataset](https://md.research.elamur.ai/free/en/papers/nabil-2015-astd-arabic-sentiment-tweets-dataset/) [Conference on Empirical Methods in Natural Language Processing (EMNLP)](https://md.research.elamur.ai/free/en/venues/emnlp/)**Mahmoud Nabil, Mohamed Aly, Amir Atiya*2015 · 470 citations*
10. **[Arabic Dialect Identification](https://md.research.elamur.ai/free/en/papers/zaidan-2014-arabic-dialect-identification/) [Computational Linguistics](https://md.research.elamur.ai/free/en/venues/cl/)**Omar F. Zaidan, Chris Callison-Burch*2014 · 437 citations*
11. **[Abusive Language Detection on Arabic Social Media](https://md.research.elamur.ai/free/en/papers/mubarak-2017-abusive-language-detection-arabic-social/) [Workshop on Abusive Language Online](https://md.research.elamur.ai/free/en/venues/workshop-on-abusive-language-online/)**Hamdy Mubarak, Kareem Darwish, Walid Magdy*2017 · 414 citations*
12. **[KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media](https://md.research.elamur.ai/free/en/papers/safaya-2020-kuisail-semeval-2020-task-12/) [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/)**Ali Safaya, Moutasem Abdullatif, Deniz Yuret*2020 · 410 citations*
13. [**The Interplay of Variant, Size, and Task Type in Arabic Pre-trained Language Models arXiv:2103.06678**Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, et al.*2021 · 365 citations*](https://md.research.elamur.ai/free/en/papers/inoue-2021-interplay-variant-size-task-type/)
14. **[Arabic Preprocessing Schemes for Statistical Machine Translation](https://md.research.elamur.ai/free/en/papers/habash-2006-arabic-preprocessing-schemes-statistical-machine/) [Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)](https://md.research.elamur.ai/free/en/venues/naacl/)**Nizar Habash, Fatiha Sadat*2006 · 359 citations*
15. **[CAMeL Tools: An Open Source Python Toolkit for Arabic Natural Language Processing](https://md.research.elamur.ai/free/en/papers/obeid-2020-camel-tools-open-source-python/) [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/)**Ossama Obeid, Nasser Zalmout, Salam Khalifa, et al.*2020 · 289 citations*
16. **[The Arabic Online Commentary Dataset: an Annotated Dataset of Informal Arabic with High Dialectal Content](https://md.research.elamur.ai/free/en/papers/zaidan-2011-arabic-online-commentary-dataset-annotated/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Omar F. Zaidan, Chris Callison-Burch*2011 · 288 citations*
17. **[Discriminating between Similar Languages and Arabic Dialect Identification: A Report on the Third DSL Shared Task](https://md.research.elamur.ai/free/en/papers/malmasi-2016-discriminating-between-similar-languages/) [International Conference on Computational Linguistics (COLING)](https://md.research.elamur.ai/free/en/venues/coling/)**Shervin Malmasi, Marcos Zampieri, Nikola Ljubešić, et al.*2016 · 287 citations*
18. **[Machine Translation of Arabic Dialects](https://md.research.elamur.ai/free/en/papers/zbib-2012-machine-translation-arabic-dialects/) [Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)](https://md.research.elamur.ai/free/en/venues/naacl/)**Rabih Zbib, Erika Malchiodi, Jacob Devlin, et al.*2012 · 286 citations*
19. **[Having Beer after Prayer? Measuring Cultural Bias in Large Language Models](https://md.research.elamur.ai/free/en/papers/naous-2023-having-beer-after-prayer-measuring/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Tarek Naous, Michael J. Ryan, Alan Ritter, et al.*2023 · 268 citations*
20. **[SemEval-2017 Task 3: Community Question Answering](https://md.research.elamur.ai/free/en/papers/nakov-2017-semeval-2017-task-3-community/) [International Workshop on Semantic Evaluation (SemEval)](https://md.research.elamur.ai/free/en/venues/semeval/)**Preslav Nakov, Doris Hoogeveen, Lluís Màrquez, et al.*2017 · 262 citations*
21. **[Morphological Analysis for Statistical Machine Translation](https://md.research.elamur.ai/free/en/papers/lee-2004-morphological-analysis-statistical-machine-translation/) [Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)](https://md.research.elamur.ai/free/en/venues/naacl/)**Young-Suk Lee*2004 · 251 citations*
22. **[Generalizing Word Lattice Translation](https://md.research.elamur.ai/free/en/papers/dyer-2008-generalizing-word-lattice-translation/) [Annual Meeting of the Association for Computational Linguistics (ACL)](https://md.research.elamur.ai/free/en/venues/acl/)**Christopher Dyer, Smaranda Muresan, Philip Resnik*2008 · 249 citations*
23. [**A review of sentiment analysis research in Arabic language arXiv:2005.12240**Oumaima Oueslati; Erik Cambria; Moez Ben HajHmida;, Habib Ounelli*2020 · 241 citations*](https://md.research.elamur.ai/free/en/papers/oueslati-2020-review-sentiment-analysis-research-arabic/)
24. **[Arabic natural language processing: An overview](https://md.research.elamur.ai/free/en/papers/guellil-2019-arabic-natural-language-processing-overview/) [Journal of King Saud University - Computer and Information Sciences](https://md.research.elamur.ai/free/en/venues/journal-of-king-saud-university-computer-and-information-sciences/)**Imane Guellil, Houda Saadane, Faical Azouaou, et al.*2019 · 239 citations*
25. **[Conventional Orthography for Dialectal Arabic](https://md.research.elamur.ai/free/en/papers/habash-2012-conventional-orthography-dialectal-arabic/) [Language Resources and Evaluation Conference (LREC)](https://md.research.elamur.ai/free/en/venues/lrec/)**Nizar Habash, Mona Diab, Owen Rambow*2012 · 237 citations*

**1** [2](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/2/) … [10](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/10/) … [16](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/16/) [next →](https://md.research.elamur.ai/free/en/topics/semitic-languages/page/2/)

Updated 24 October 2026

- [← previous topic Semantic role labeling](https://md.research.elamur.ai/free/en/topics/semantic-role-labeling/)
- [next topic → Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)