This dataset contains only bibliographic metadata (titles, abstracts, author names, affiliations, DOIs, citation counts, etc.) aggregated from publicly available sources. The copyright and intellectual property rights to the original scholarly texts (including the full text of papers) belong to their respective authors, publishers, or institutions.

ArabicNLPWorld does not claim ownership over the content of the papers, nor does it license the original texts under CC BY-SA 4.0. The CC BY-SA 4.0 license applies exclusively to:

The dataset's structure and format (e.g., the organisation of fields, the Parquet schema);
The curated annotations (e.g., automatically assigned topic_name, matched_terms, concepts, and the pedagogy-filtering logic);
The aggregation and deduplication work.

The abstracts and titles are reproduced as they appear in the original sources and remain under the copyright of their original rights holders. Their inclusion is considered fair use for academic research and bibliographic indexing purposes.

If you are a rights holder and believe that your work is included in a way that infringes your rights, please contact us via the Hugging Face repository issues, and we will promptly remove the offending metadata under a notice‑and‑takedown procedure.

This practice is consistent with established data curation efforts (e.g., HPLT, Common Crawl, etc.), where the data packaging and derived annotations are licensed openly, while the underlying content is not.

Dataset Details

Dataset Description

The Arabic NLP Corpus is a curated collection of 7,143 scholarly papers focused on Arabic Natural Language Processing (NLP). It compiles metadata from six major academic sources: arXiv, ACL Anthology, Semantic Scholar, Crossref, and OpenAlex (with two separate collection runs). The dataset is designed to support bibliometric analysis, literature reviews, trend detection, and benchmarking in Arabic NLP research.

Each record includes full bibliographic metadata, citation counts, author affiliations, geographical information, and automatically extracted topic labels. The corpus has been filtered to retain only papers with abstracts and to remove pedagogy‑focused works (using a rule‑based filter with NLP‑specific protection). The result is a high‑quality, research‑ready dataset for both quantitative and qualitative studies of the Arabic NLP landscape.

Curated by: ArabicNLPWorld (Hugging Face organization)
Funded by: Not applicable (community effort)
Shared by: ArabicNLPWorld
Language(s): Primarily English abstracts; Arabic keywords and dialect terms are present in titles/abstracts.
License: CC BY-SA 4.0 (for the dataset packaging and annotations only – see Legal Notice above).

Dataset Sources

Repository: https://huggingface.co/datasets/ArabicNLPWorld/arabic-nlp-corpus
Analysis Report: A comprehensive analysis (topic modeling, citation trends, dialect distribution, co‑authorship networks, regression, etc.) is included in the repository under analysis_report.txt.

Uses

Direct Use

This dataset is intended for:

Bibliometric analysis – studying publication trends, citation patterns, and research impact.
Topic modeling and trend detection – identifying emerging subfields and shifts in research focus.
Author and institutional network analysis – mapping collaborations and influence.
Dialect and task analysis – investigating which dialects and NLP tasks receive the most attention.
Training and evaluation of NLP models – e.g., for citation prediction, topic classification, or author disambiguation (using the metadata as features).

Out-of-Scope Use

The dataset does not contain full‑text PDFs; it only includes metadata and abstracts. It should not be used to reproduce the full content of papers.
It is not intended for generating summaries or translations of the papers themselves.
The dataset is not a gold standard for any NLP task; it is a bibliographic corpus, and its use for training models that make high‑stakes decisions is discouraged without additional validation.

Dataset Structure

The dataset is stored in Parquet format (Snappy compression) and is organized as follows:

data/
└── master_arabic_nlp_clean_filtered.parquet

Each row in the Parquet file corresponds to one paper and contains the following fields:

Field	Type	Description
`id`	string	Unique identifier (source‑specific)
`title`	string	Paper title
`abstract`	string	Paper abstract
`authors`	list[string]	List of author names
`year`	string	Publication year (4‑digit)
`venue`	string	Conference/journal name
`doi`	string	Digital Object Identifier
`pdf_url`	string	Link to PDF (if available)
`source`	string	Original source(s), concatenated with `;` if multiple
`countries`	list[string]	Affiliation countries (ISO codes)
`institutions`	list[string]	Affiliation institutions
`concepts`	list[string]	OpenAlex concepts (when available)
`citation_count`	int	Number of citations
`matched_terms`	list[string]	Arabic‑related keywords found in title/abstract
`relevance_score`	float	Relevance score (if computed; currently 0 for all)
`topic_name`	string	BERTopic‑assigned topic label

All list fields are stored as native Parquet lists for efficient querying.

Dataset Creation

Curation Rationale

Arabic NLP has grown rapidly, but there was no unified, open‑access bibliographic dataset covering the entire field. This corpus was created to fill that gap, providing a solid foundation for systematic reviews, trend analysis, and future benchmarking. The curation process combined multiple open‑access metadata sources and applied careful deduplication and filtering to ensure high quality.

Source Data

The data was sourced from:

arXiv – preprints (filtered for Arabic NLP relevance) – 613 papers.
ACL Anthology – official proceedings of ACL, EMNLP, and affiliated workshops – 657 papers.
Semantic Scholar – broad academic graph with citation data – 6,865 papers.
Crossref – DOI registration metadata – 3,985 papers.
OpenAlex – comprehensive scholarly database (two separate collection runs) – 4,054 and 206 papers, respectively.

All sources are publicly accessible and provide metadata under open terms.

Data Collection and Processing

Query‑based retrieval – We used 60+ search queries (e.g., "arabic dialect", "AraBERT", "Arabic sentiment analysis") to retrieve candidate papers from each source.
Deduplication – Records were deduplicated by DOI and then by title + year, merging metadata from multiple sources.
Relevance filtering – Papers were kept only if they had at least two Arabic‑related matched_terms (e.g., "arabic", "dialect", "translation").
Pedagogy removal – Papers containing strong pedagogical phrases (e.g., "teaching", "classroom", "curriculum") were removed, unless they also contained NLP‑specific indicators (e.g., "transformer", "BERT", "sentiment analysis") to avoid removing technical works.
Abstract requirement – Only papers with a non‑empty abstract of at least 10 characters were retained.
Year filtering – Papers with publication years before 1960 were discarded.

Who are the source data producers?

The original data producers are the authors, journals, conferences, and preprint servers that published the papers. The metadata was aggregated and processed by the ArabicNLPWorld team.

Annotations

No manual annotations were added. The matched_terms and topic_name were computed automatically using rule‑based keyword matching and BERTopic (with a multilingual embedding model). The concepts field comes from OpenAlex's automatic concept tagging.

Personal and Sensitive Information

This dataset contains only bibliographic metadata, which includes author names and affiliations. It does not contain any personally identifiable information beyond academic affiliations, which are publicly available in the original sources. No effort was made to anonymize the data because the information is already public and aggregated at the paper level.

Bias, Risks, and Limitations

Coverage bias – The dataset is biased towards sources and queries used; it may under‑represent papers published in non‑English venues, closed‑access journals, or regional conferences not indexed by the chosen APIs.
Language bias – While the focus is Arabic NLP, most abstracts are in English; papers published primarily in Arabic may be underrepresented.
Temporal bias – The collection is snapshotted in 2026; more recent papers may be missing, and the citation counts are frozen at the time of collection.
Topic modeling artifacts – The BERTopic labels are generated automatically and may not perfectly reflect human‑defined categories. The dominant topic (0_text_speech_translation_recognition) is very broad and may obscure finer sub‑areas.
Filtering limitations – The pedagogy removal filter, although protected by NLP indicators, may still have incorrectly excluded some relevant technical papers or kept some pedagogical ones.
Citation data – Citation counts come from different sources and may be inconsistent (e.g., Semantic Scholar vs. Crossref).

Recommendations

Users should:

Interpret the dataset as a research tool rather than a definitive ground truth.
Consider supplementing with other sources (e.g., IEEE, Springer) for a more comprehensive view.
Be cautious when generalizing trends to the entire field, as the corpus is not exhaustive.
If using the dataset for training models, be aware of the biases and possibly apply domain adaptation.

Analysis Highlights

From the accompanying analysis (analysis_report.txt), key findings include:

Publication growth – A sharp increase after 2018 (AraBERT) and an explosion in 2023 (ChatGPT/LLM boom).
Top topics – The largest topic is 0_text_speech_translation_recognition (2,955 papers), followed by sentiment analysis, medical/psychological questionnaires, dialectology, NER, and Quranic studies.
Most cited papers – Dominated by foundational models (AraBERT, ARBERT/MARBERT) and classic linguistic works.
Top authors – Nizar Habash (151 papers), Muhammad Abdul-Mageed (95), Mona Diab (86), Wajdi Zaghouani (57), Ahmed Ali (55), etc.
Top institutions – King Saud University, Cairo University, Columbia University, University of Jordan, King Abdulaziz University.
Dialect distribution – MSA (Modern Standard Arabic) is mentioned in 1,553 papers, Egyptian in 462, Maghrebi in 341, Levantine in 187, and Gulf in 105.
Task–dialect gaps – Zero publications exist for many rare dialect–task combinations (e.g., Hejazi with sentiment analysis, Hassaniya with NER, etc.), highlighting future research opportunities.
Citation drivers – Regression analysis shows that papers from OpenAlex and Semantic Scholar, with more matched_terms, and with institutional affiliations receive significantly more citations.

Citation

If you use this dataset in your research, please cite it as:

BibTeX:

@misc{arabic_nlp_corpus,
  author = {{ArabicNLPWorld Community}},
  title = {Arabic NLP Corpus: A Curated Collection of Scholarly Papers},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ArabicNLPWorld/arabic-nlp-corpus}
}

APA:

ArabicNLPWorld Community. (2026). Arabic NLP Corpus: A Curated Collection of Scholarly Papers. Hugging Face. https://huggingface.co/datasets/ArabicNLPWorld/arabic-nlp-corpus

Glossary

BERTopic – A topic modeling technique that uses BERT embeddings and clustering to generate coherent topics.
Matched terms – Keywords from a predefined list (e.g., "arabic", "dialect", "translation") that appear in a paper's title or abstract. Papers with at least two such terms were retained.
OpenAlex – A large open bibliographic database that provides metadata on scholarly works, authors, institutions, and concepts.
MSA – Modern Standard Arabic, the formal literary standard used in writing and formal speech across the Arab world.

More Information

A full analysis report, including detailed statistics, topic trends, dialect distributions, co‑authorship networks, geographic maps, regression results, and task–dialect gap matrices, is available in the repository (see analysis_report.txt). The report also contains χ² tests and topic growth rates (CAGR).

Dataset Card Authors

This dataset card was prepared by the ArabicNLPWorld team.

Dataset Card Contact

For questions or suggestions, please open an issue on the Hugging Face repository or contact the organization maintainers. ```

Downloads last month: -

Total file size:

6.6 MB