AraBERT fine-tuned for Arabic NER (ANERCorp)
Fine-tune of aubmindlab/bert-base-arabertv02 for Arabic Named Entity Recognition on ANERCorp. Tags tokens with four entity types using the BIO scheme: PERS, LOC, ORG, MISC.
Results (ANERCorp test split, entity-level / span-level)
| Entity | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| LOC | 0.8769 | 0.9275 | 0.9015 | 676 |
| PERS | 0.8640 | 0.8335 | 0.8485 | 907 |
| ORG | 0.7130 | 0.7037 | 0.7083 | 459 |
| MISC | 0.6571 | 0.5679 | 0.6093 | 243 |
| micro | 0.8185 | 0.8070 | 0.8127 | 2285 |
F1 is computed at the entity (span) level: a prediction counts as correct only when the full entity span and its type match the gold annotation.
Training
- Base:
aubmindlab/bert-base-arabertv02 - 3 epochs, lr 2e-5, batch size 16, weight decay 0.01
- Sub-word alignment: the first sub-token of each word keeps the label; the remaining sub-tokens are set to
-100and ignored by the loss. - Sentences segmented on
.!?before tokenization to preserve context.
Usage
from transformers import pipeline
ner = pipeline("token-classification", model="Komail262/arabert-ner-anercorp", aggregation_strategy="simple")
ner("أعلن المدير التنفيذي لشركة أبل تيم كوك عن افتتاح فرع جديد في الرياض.")
License & attribution
This model is a fine-tune of aubmindlab/bert-base-arabertv02 (AraBERT) on the ANERCorp dataset. Rights to the base model and dataset remain with their original authors; please consult their pages for licensing terms.
Limitations
- Trained on ANERCorp (news-domain Modern Standard Arabic); performance on dialectal or social-media Arabic is expected to be lower.
- MISC and ORG lag behind LOC and PERS. MISC is a heterogeneous category with the least training support.
- Not intended for high-stakes use without domain-specific evaluation.