--- title: README emoji: đ colorFrom: indigo colorTo: indigo sdk: static pinned: false --- # ArabicNLPWorld â Arabic MSA, Dialects & LowâResource NLP Research Hub       **ArabicNLPWorld** is a research organization dedicated to natural language processing for **Modern Standard Arabic (MSA)** â a wellâresourced language â as well as **underâresourced Arabic dialects**, **lowâresource language pairs involving Arabic**, **Islamic religious texts**, and **ArabicâRussian translation**. We develop and share openâsource models, datasets, and educational tools to bridge the digital divide across all varieties and modalities of Arabic. > đ **This is an organization card.** Our models, datasets, and demos are available on our [Hugging Face Organization Page](https://huggingface.co/ArabicNLPWorld). --- ## đŻ Our Mission - Build stateâofâtheâart language models for **Modern Standard Arabic (MSA)** â leveraging its rich existing resources. - Create resources and models for **underâresourced Arabic dialects** (Egyptian, Levantine, Gulf, Maghrebi, Sudanese, etc.). - Advance **ArabicâRussian machine translation** using our 15.8M parallel corpus. - Support **lowâresource language pairs** where Arabic is one side (e.g., Arabic â Tatar, Arabic â Chechen, Arabic â Bashkir, Arabic â Hausa, Arabic â Somali). - Develop specialised NLP tools for **Islamic religious texts**: - **The Quran** with Russian translation (Elmir Kuliev) - **Sahih al-Bukhari** â the most authentic hadith collection - **Sahih Muslim** â the second most authentic collection - **40 Hadith of al-Nawawi** (41 in some editions) - **Kutub al-Sittah (The Six Major Hadith Collections)** â including Sunan Abu Dawud, Jami` at-Tirmidhi, Sunan an-Nasa'i, and Sunan Ibn Majah - Foster a community of researchers, developers, native speakers, dialect speakers, and Islamic scholars working together on inclusive Arabic NLP. --- ## đ§ Clarification: MSA vs. Dialects vs. LowâResource | Variety / Pair | Resource Status | Description | |----------------|----------------|-------------| | **Modern Standard Arabic (MSA)** | â **Wellâresourced** | Hundreds of billions of tokens, many pretrained models (AraBERT, MARBERT, AraT5, CAMeLBERT), large parallel corpora with English and other major languages. | | **Arabic dialects** (Egyptian, Levantine, Gulf, Maghrebi, etc.) | â ď¸ **Underâresourced to lowâresource** | Limited annotated data, few pretrained models, scarce parallel corpora with MSA or English. Egyptian is bestâresourced among dialects but still far behind MSA. | | **Arabic â Russian translation** | đ **Midâresource** | Our 15.8M corpus is the largest publicly available for this pair, but still modest compared to EnglishâArabic (100M+). | | **Lowâresource pairs** (Arabic â Turkic, Caucasian, African languages) | â **Lowâresource** | Very few (often zero) parallel datasets; requires transfer learning, data augmentation, and zeroâshot techniques. | | **Islamic religious texts** | đ **Domainâspecific** | Rich but specialised vocabulary (classical Arabic). Includes **Quran**, **Sahih al-Bukhari**, **Sahih Muslim**, **40 Hadith of al-Nawawi**, and **Kutub al-Sittah** with curated parallel translations. | --- ## đ Interactive Demos Explore our live Hugging Face Spaces and try out our models directly in your browser: ### **đ¤ Language Models** - **[AraBERT Playground]()** â Generate and analyze MSA text. - **[DialectBERT Explorer]()** â Pretrained model for Egyptian, Levantine, and Gulf Arabic. - **[ArabicâRussian Embeddings]()** â Crossâlingual word vectors for translation. ### **đ Machine Translation** - **[Arabic â Russian Translator]()** â Neural translation demo (15.8M parallel pairs). - **[MSA â Dialect Translator]()** â Convert between Modern Standard Arabic and Egyptian/Levantine. - **[Quran & Hadith Translation Explorer]()** â Arabic originals with Russian (Kuliev) and English parallels. ### **đ Linguistic Tools** - **[Arabic Morphological Analyzer]()** â Rootâbased segmentation and POS tagging. - **[Dialect Identifier]()** â Detect MSA vs. Egyptian, Levantine, Gulf, Maghrebi. - **[Named Entity Recognition for Arabic]()** â Identify persons, locations, organizations. ### **đ Data & Benchmarks** - **[ArabicâRussian Corpus Explorer]()** â Browse 15.8M parallel sentences. - **[Dialect NLP Leaderboard]()** â Compare model performance on dialect tasks. - **[Islamic Text Annotation Tool]()** â Help us improve Quran/hadith alignments. *Click on any demo to start experimenting â no installation required!* --- ## đ§ Research Focus Areas ### **đ¸đŚ Modern Standard Arabic (MSA) â WellâResourced** - Continued pretraining and fineâtuning of MSA models (AraBERT, AraT5, MARBERT) - Benchmarking on standard tasks (POS, NER, sentiment, QA) - Leveraging MSA as a source for transfer learning to dialects ### **đŁď¸ Arabic Dialects â UnderâResourced to LowâResource** Focus on: Egyptian (arz), Levantine (apc), Gulf (afb), Maghrebi (ary), Sudanese (apd) **Challenges we address:** - Lack of annotated data â data augmentation, semiâsupervised learning - Few parallel corpora (dialect â MSA, dialect â English) - Absence of dialectâspecific pretrained models **Our approach:** - Crossâlingual transfer from MSA to dialects - Fewâshot and zeroâshot learning for dialect tasks - Crowdsourced annotation and validation with native speakers ### **đ ArabicâRussian Bilingual NLP â MidâResource** - **15,801,992 parallel sentences** (our flagship corpus) - Sources: OPUS, TED, Baranov dictionary, Borisov dictionary, Sahih al-Bukhari, Sahih Muslim, 40 Hadith, Quran (Kuliev), phrasebook, Tatoeba - Length correlation: 0.925 - Applications: translation, crossâlingual retrieval, bilingual lexicography ### **đ LowâResource Pairs Involving Arabic â LowâResource** We focus on language pairs with minimal or no parallel data: | Pair | Resource Status | Our Work | |------|----------------|----------| | Arabic â Tatar | Very low | Data collection, transfer learning from ArabicâRussian + RussianâTatar | | Arabic â Chechen | Extremely low | Zeroâshot translation via English or Russian pivot | | Arabic â Bashkir | Extremely low | Crossâlingual embeddings | | Arabic â Hausa | Very low | Leveraging NLLB model | | Arabic â Somali | Very low | Data collection and annotation | ### **đ Islamic Religious Texts â DomainâSpecific** We provide digitised, aligned, and machineâreadable versions of major Islamic texts: | Text | Description | Parallel Translation | |------|-------------|----------------------| | **The Quran** | The holy book of Islam, 114 surahs | Russian (Elmir Kuliev), English (Sahih International) | | **Sahih al-Bukhari** | Most authentic hadith collection (c. 7,000+ hadith) | Russian translation | | **Sahih Muslim** | Second most authentic collection (c. 7,000+ hadith) | Russian translation | | **40 Hadith of al-Nawawi** | Concise collection of 40 (or 41) essential hadith | Russian translation | | **Sunan Abu Dawud** | One of the six major collections (Kutub al-Sittah) | Russian (in progress) | | **Jami` at-Tirmidhi** | One of the six major collections | Russian (in progress) | | **Sunan an-Nasa'i** | One of the six major collections | Russian (in progress) | | **Sunan Ibn Majah** | One of the six major collections | Russian (in progress) | **Applications:** - Semantic search over hadith corpora - Question answering on Islamic texts - Classical Arabic morphological analysis - Crossâcollection hadith matching (e.g., finding the same hadith in Bukhari and Muslim) - Alignment of multiple translations for linguistic study ### **đ Lexicographic Resources** - **ArabicâRussian Dictionary** â Kh.K. Baranov (latest edition) â digitised and aligned - **RussianâArabic Dictionary** â V.M. Borisov (latest edition) â bidirectional coverage - Machineâreadable formats for NLP integration --- ## đ Educational Resources We believe in **open education** and **reproducible research**. All our tutorials and teaching materials are freely available. - **[Interactive Notebooks]()** â Arabic NLP, dialect processing, ArabicâRussian MT, lowâresource techniques (in Python, using Hugging Face libraries) - **[Video Lectures]()** â Recorded talks on Arabic morphology, dialect identification, and Islamic text processing - **[Course Materials]()** â Slides, readings, and assignments from our university courses - **[Blog Posts]()** â Deep dives into challenges and solutions for Arabic dialects and lowâresource pairs --- ## đ¤ Get Involved We welcome contributions from the community â researchers, developers, students, native speakers, dialect speakers, and Islamic scholars. ### **For Researchers** - Use our models and datasets (and cite us!) - Collaborate on dialect annotation or lowâresource pair projects - Contribute new benchmarks for dialects or ArabicâRussian MT ### **For Developers** - Integrate our models into translation, search, or chatbot applications - Report bugs or suggest improvements via GitHub Issues - Submit pull requests to our openâsource repositories ### **For Native & Dialect Speakers** - Help us validate dialect annotations and translations - Share dialect texts (with permission) to enrich our data - Provide feedback on model outputs to reduce errors ### **For Islamic Scholars & Students** - Help verify Quranic verse alignments and hadith translations - Suggest improvements for religious text processing - Use our tools for digital Islamic studies ### **For Students** - Use our demos and tutorials for learning - Participate in our mentorship program or summer schools - Start your own research project with our support --- ## đ Corpus Highlights Our flagship resource â the **ArabicâRussian Translation Corpus**: | Statistic | Value | |-----------|-------| | Total pairs | 15,801,992 | | Length correlation | 0.925 | | Arabic tokens | 357.7M | | Russian tokens | 366.0M | | Unique Arabic tokens | 1,848,317 | | Unique Russian tokens | 933,467 | | Sources | OPUS, TED, Baranov, Borisov, **Sahih al-Bukhari**, **Sahih Muslim**, 40 Hadith, Quran (Kuliev), phrasebook, Tatoeba | **Most frequent Arabic words:** ŮŮ (13.68M), Ů Ů (8.45M), ŘšŮŮ (5.59M) **Most frequent Russian words:** и (15.88M), в (15.52M), пО (5.38M) --- ## đ Connect With Us - **đ¤ Hugging Face**: [ArabicNLPWorld](https://huggingface.co/ArabicNLPWorld) â Models, datasets, and spaces - **đ§ Contact**: arabicnlpworld@example.com --- ## đ Ecosystem Integration Our work is integrated with the broader Hugging Face ecosystem: - **Models** on the Hub with easyâtoâuse pipelines - **Datasets** with streaming and evaluation scripts - **Spaces** for interactive demos and educational tools - **Gradio** apps for userâfriendly interfaces --- **Empowering Arabic MSA, dialects, lowâresource pairs, and Islamic texts through open science and community collaboration.**