hbku/arabic-gov-ner
Arabic Gov NER
Macro F1
0.89
F1 (PERSON)
0.92
F1 (LAW_REF)
0.84
QID redaction recall
0.97
About this model
Token-classification model that extracts persons, organisations, law and decree references, dates, and QID mentions from Arabic government documents. Ships with a redaction mode that masks QIDs and personal names for safe document sharing.
Intended use
Entity extraction and PII redaction in document-processing pipelines for circulars, decisions, and correspondence across government entities.
Training data lineage
Fine-tuned on 62,000 annotated sentences drawn from the arabic-gov-docs-corpus, with QID patterns validated against qid-specimen-templates.
Limitations & bias notes
Recall on transliterated foreign organisation names is noticeably lower than on Arabic-native entities. Law references using informal shorthand (e.g. omitting the year) are frequently missed or partially tagged.
Evaluation metrics
| Macro F1 | 0.89 |
| F1 (PERSON) | 0.92 |
| F1 (LAW_REF) | 0.84 |
| QID redaction recall | 0.97 |
Try it
Live sandboxFeedback
Owning entity
- Updated
- 2026-05-06
- Latest version
- 1.4.0
- License
- Open
- Access
- Open