ELNER-DZ: Algerian Arabic Dataset for Named Entity Recognition and Entity Linking

📌 Official DOI:
🤗 Also available on Hugging Face
📥 Download Dataset (.rar from Zenodo)

This dataset, titled ELNER-DZ, was created by Bouguettoucha Hadjer Hanine and Djouablia Ilhem as part of our Master’s thesis. It is the first large-scale dataset designed for Named Entity Recognition (NER) and Entity Linking (EL) in Algerian Arabic Dialect (Darija), including both Arabic script and Arabizi (Latin-script).

This dataset contains over 2 million dialectal sentences labeled with more than 1.9 million named entities and linked to Wikidata QIDs.

🧾 Dataset Summary

• Name: ELNER-DZ
  
• Languages: Arabic (ar for MSA, arq for dialectal), Arabizi (Latin), French (fr), English (en)
  
• Script: Arabic and Latin (Arabizi)
  
• Format: JSON (compressed in data.rar)
  
• Annotations:
  • Named Entity spans (start, end)
  • NER labels (PER, LOC, ORG, etc.)
  • Normalized forms
  • Wikidata QIDs

📁 File Structure

• data/data.rar — Compressed archive containing data.json
• examples/loading_example.py — Script to extract and load the dataset
• LICENSE — CC-BY-4.0
• dataset_card.md — Hugging Face dataset summary

✨ Example Format

json { "id": 188, "text": "3reft wa7ed lperson khadem f Yassir", "entities": [ { "start": 29, "end": 35, "label": "ORG", "wikidata_id": "Q117156470", "normalized": "Yassir" } ] } `

🏷️ Entity Types

• PER: Person
• LOC: Location
• ORG: Organization
• PROD: Product
• LAW: Legal texts or rules
• LANG: Language
• EVENT: Events
• DATE: Temporal expressions
• NORP: Nationality/Religious/Political groups
• SPORT: Sports & Competitions
• SYMPTOM, DISEASE: Medical categories
• MISC: Miscellaneous

🧪 Tasks Supported

• Named Entity Recognition (NER)
• Entity Linking (EL) with Wikidata
• Dialectal NLP in Algerian Arabic
• Code-switching and multiscript modeling
• Low-resource transfer learning

🧰 How to Use

▶️ Requirements

bash pip install datasets rarfile sudo apt-get install unrar # For Linux

▶️ Run the loading script

bash python examples/loading_example.py

Or manually extract and load:

```python import rarfile rf = rarfile.RarFile("data/data.rar") rf.extractall("data/")

from datasets import loaddataset dataset = loaddataset("json", data_files="data/data.json", split="train") print(dataset[0]) ```

🔍 Dataset Details

• Source: Social media, dialogues, e-commerce, Wikidata SPARQL

• Annotation:

  • Semi-automated and rule-based extraction
  • Manual normalization of entity surface forms
  • Wikidata QID linking via SPARQL and fallback search

👩‍💻 Authors

• Bouguettoucha Hadjer Hanine
• Djouablia Ilhem

📄 License

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

🔗 View Full License

📚 Citation

bibtex @dataset{bouguettoucha_djouablia_2025, author = {Bouguettoucha, Hadjer Hanine and Djouablia, Ilhem}, title = {ELNER-DZ: A Dataset for Named Entity Recognition and Linking in Algerian Arabic}, year = 2025, publisher = {Zenodo}, doi = {10.5281/zenodo.15798592}, url = {https://doi.org/10.5281/zenodo.15798592} }