AQMAR Arabic NER – Full Label Set
Dataset Summary
This dataset is a reprocessed version of the AQMAR Arabic Named Entity Recognition (NER) corpus originally released by the Arabic NLP group at Carnegie Mellon University (CMU).
It preserves the full original AQMAR annotation scheme, including fine-grained miscellaneous entity categories (MIS0, MIS1, MIS2, etc.).
This version is intended for linguistic analysis and annotation studies, not as a clean leaderboard benchmark.
Source
Original data obtained from: https://www.cs.cmu.edu/~ark/ArabicNER/
The original files are in CoNLL format:
- One token per line
- Empty lines indicate sentence boundaries
Data Processing
- Sentence boundaries are taken directly from the original files
- No re-tokenization was applied
- File-level train/dev/test split (20/4/4 files) to avoid sentence leakage
Label Set
Includes:
- PER, LOC, ORG
- Fine-grained MIS* categories
- O (outside)
Some fine-grained labels are rare and may not appear in dev/test. Macro-F1 over all labels is therefore not recommended.
Intended Use
Recommended:
- Fine-grained NER analysis
- Annotation ontology studies
- Error analysis
Not recommended:
- Benchmark comparisons using macro-F1 over all labels