Arabic Named Entity Recognition (NER) using BERT-Arabic
This model is a fine-tuned version of asafaya/bert-base-arabic for Arabic Named Entity Recognition (NER) on the Arabic subset of the PAN-X (XTREME) dataset.
It predicts three primary entity types: Persons (PER), Organizations (ORG), and Locations (LOC), using the standard IOB2 tagging scheme.
📊 Model Performance & Test Results
The model was evaluated on an unseen test set split (10% of total stratified data).
Overall Metrics
| Metric | Score | Percentage |
|---|---|---|
| Test F1-Score (Entity-level) | 0.9172 | 91.72% |
| Test Accuracy (Token-level) | 0.9634 | 96.34% |
| Token Error Rate | 0.0366 | 3.66% |
| Entity Error Rate | 0.0828 | 8.28% |
Detailed Classification Report
precision recall f1-score support
LOC 0.93 0.93 0.93 1564
ORG 0.88 0.89 0.89 1438
PER 0.93 0.94 0.94 1541
micro avg 0.91 0.92 0.92 4543
macro avg 0.91 0.92 0.92 4543
weighted avg 0.91 0.92 0.92 4543