MAGHREB-HOF: A Large-Scale Annotated Corpus for Hate and Offensive Language Detection in Maghrebi Arabic
Year: 2022
Type: dataset
Abstract: MAGHREB-HOF: A Large-Scale Corpus for Hate and Offensive Language Detection in Maghrebi Arabic MAGHREB-HOF is a large-scale annotated corpus for hate and offensive language (HOF) detection in Maghrebi Arabic social-media text. The dataset contains 200,000 anonymized comments and supports both binary and fine-grained classification. The corpus includes dialectal Arabic, Arabizi, and code-mixed Arabic–French/English content commonly observed in North African online communication. It is intended to support reproducible research on hate and offensive language detection in noisy dialectal environments characterized by spelling variation, dialect mixing, transliteration, and informal writing. The dataset accompanies the peer-reviewed article: Ahmed Zoubir Messalti and Sadik Bessou, “Word and Character Models for Maghrebi Hate Speech Detection: A Large-Scale Facebook Corpus,” International Journal of Computers and their Applications, vol. 33, no. 3, pp. 255–266, 2026. The article is indexed in Scopus. What This Record Contains This Zenodo record provides: The annotated corpus of anonymized Maghrebi social-media comments. The annotation guidelines used during the labeling process. Label definitions and metadata supporting reproducible experimentation. The official split-index file used in the experiments (.xlsx). A README file describing the official experimental partitions. The provided partition information supports reproduction of the two experimental tracks reported in the companion paper: Sparse linear models: stratified 70/30 holdout split and five-fold cross-validation where applicable. Neural sequence models: stratified 70/15/15 train/validation/test split. Researchers are encouraged to use the provided partitions when directly comparing new models with the results reported in the companion article. Download and Use The dataset can be downloaded from the Files section of this Zenodo record when access to the files is available. After downloading: Extract the dataset files. Consult the README and annotation guidelines before preprocessing the corpus. Select the required task formulation: Binary classification: Clean vs. Hate/Offensive Language. Fine-grained classification: Clean, Hate, Insult, and Profanity. Use the provided split-index file when reproducing the experiments reported in the companion article. Clearly document any alternative preprocessing procedure, data partition, label transformation, or filtering applied to the original dataset. For reproducibility, publications using MAGHREB-HOF should report at least: task formulation; preprocessing procedure; train/validation/test partition; model configuration; random seed where applicable; evaluation metrics. Source and Data Collection The corpus was constructed from comments collected from public Facebook pages and posts associated with the Maghrebi region. The collected discussions cover topics that may generate substantial online interaction and potentially hostile exchanges, including politics, identity-related issues, social topics, and sports rivalries. All released comments are anonymized. Direct personal identifiers are not included in the distributed dataset. Annotation and Labels Each comment is annotated for hate and offensive language. Binary Classification 0 — Clean / non-offensive 1 — Hate or offensive content Fine-Grained Classification Clean Hate Insult Profanity The annotation guidelines included in this record provide the definitions and decision rules used during corpus construction. Core Columns Typical dataset fields include: ID — unique internal comment identifier Comments_origine — original comment text label — binary HOF label (0 / 1) label_multi — fine-grained label (Clean, Hate, Insult, or Profanity) Comments_Date Annotation_Date Page Post Hub Post_Date City — predicted dialect/city metadata Score — prediction confidence score associated with the dialect-identification model The City and Score fields were generated using the pretrained Hugging Face model: https://huggingface.co/Ammar-alhaj-ali/arabic-MARBERT-dialect-identification-city These predictions are included as automatically generated descriptive metadata. They are not manually validated dialect annotations and should therefore not be treated as gold-standard geographic or dialect labels. Recommended Use MAGHREB-HOF can support research on: hate speech detection; offensive language detection; toxicity detection; Maghrebi Arabic NLP; dialect-robust text classification; Arabizi processing; Arabic–French/English code-mixed NLP; character-, word-, and subword-level text representation; content moderation. The dataset may also be used to evaluate robustness to spelling variation, informal writing, transliteration, and lexical variation in Maghrebi social-media text. Citation If you use MAGHREB-HOF, in whole or in part, in a scientific publication, thesis, dissertation, software project, benchmark, or other research output, please cite both the dataset record and the companion peer-reviewed article. Companion Article Ahmed Zoubir Messalti and Sadik Bessou, “Word and Character Models for Maghrebi Hate Speech Detection: A Large-Scale Facebook Corpus,” International Journal of Computers and their Applications, vol. 33, no. 3, pp. 255–266, 2026. BibTeX — Companion Article @article{Messalti2026WordCharacter, author = {Messalti, Ahmed Zoubir and Bessou, Sadik}, title = {Word and Character Models for Maghrebi Hate Speech Detection: A Large-Scale Facebook Corpus}, journal = {International Journal of Computers and their Applications}, year = {2026}, volume = {33}, number = {3}, pages = {255--266}, issn = {1076-5204}, publisher = {International Society for Computers and Their Applications} } Dataset Ahmed Zoubir Messalti and Sadik Bessou, MAGHREB-HOF: A Large-Scale Corpus for Hate and Offensive Language Detection in Maghrebi Arabic, Zenodo, 2026. https://doi.org/10.5281/zenodo.19048125 BibTeX — Dataset @dataset{Messalti2026MAGHREBHOF, author = {Messalti, Ahmed Zoubir and Bessou, Sadik}, title = {MAGHREB-HOF: A Large-Scale Corpus for Hate and Offensive Language Detection in Maghrebi Arabic}, year = {2026}, publisher = {Zenodo}, doi = {10.5281/zenodo.19048125}, url = {https://doi.org/10.5281/zenodo.19048125} } Ethical Considerations The corpus contains potentially harmful and offensive language, including insults, profanity, and hate speech. It is released for scientific research aimed at understanding, detecting, and mitigating harmful online content. All distributed comments are anonymized and direct personal identifiers have been removed. Researchers using the dataset remain responsible for complying with applicable institutional, legal, and ethical requirements concerning social-media data and potentially harmful content. The corpus should not be used to identify, profile, or target individual users. License Creative Commons Attribution 4.0 International (CC BY 4.0). Users may reuse and adapt the dataset in accordance with the license conditions, provided appropriate attribution is given. Scientific works using MAGHREB-HOF should cite both: this Zenodo dataset record; and the companion peer-reviewed article. Access Dataset files are distributed through this Zenodo record according to the access conditions indicated on the record page. If an embargo is active, the metadata and DOI remain publicly visible while access to the dataset files is restricted until the embargo expires. Related Publication Ahmed Zoubir Messalti and Sadik Bessou“Word and Character Models for Maghrebi Hate Speech Detection: A Large-Scale Facebook Corpus.”International Journal of Computers and their Applications, vol. 33, no. 3, pp. 255–266, 2026. Dataset Statistics Dashboard An interactive dashboard providing descriptive statistics, label distributions, metadata summaries, and exploratory views is available at: http://magh-arabic-hof.duckdns.org:3000/public/dashboard/4a359fc9-bfe5-498b-93f6-957335e5ccf1/page The dashboard is provided for exploration and reporting purposes and does not provide access to restricted or embargoed dataset files. Annotation Demo A Doccano instance demonstrating the annotation interface and workflow is available at: http://magh-arabic-hof.duckdns.org:8800/auth Access credentials for the demonstration environment may be provided privately to reviewers or researchers when appropriate. Contact Ahmed Zoubir MessaltiLRSD LaboratoryDepartment of Computer Science, Faculty of SciencesFerhat Abbas Sétif 1 UniversityAlgeria Email: [email protected] Keywords Maghrebi Arabic; Arabizi; code-mixing; hate speech; offensive language; toxicity detection; Arabic NLP; social media corpus; machine learning; deep learning; text classification; content moderation
Authors: BESSOU, SADIK, MESSALTI, Ahmed Zoubir
Institution: University Ferhat Abbas of Setif
Language: arb
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7137142000