AchrafSoltani/morisco-ner

Darija Arabizi NER

A native-validated NER test set for Moroccan Darija written in Arabizi (Latin script, with digits such as 3, 7, 9), four entity types (PER, LOC, ORG, MISC), plus the transliterated training data used in our experiments. To our knowledge this is the first native-validated Arabizi Darija NER test set.

Contents

• arabizi_gold.jsonl — the gold test set: 400 sentences, 252 entities, native-validated. Format: {"tokens": [...], "spans": [[start, end, type], ...]} (token indices, inclusive).
• arabizi_draft_review.csv — annotation workbook: LLM-teacher draft (predicted_entities) + native corrections (CORRECTION). Released for transparency / error analysis.
• arabizi_gold_ar.jsonl — the gold transliterated to Arabic script (token-aligned), for transliterate-then-NER baselines.
• darnercorp_gs_big_arabizi.jsonl — transliterated training data (DarNERcorp gold + LLM-teacher silver → Arabizi).

Provenance & attribution

• Arabizi sentences derived from DODa — Darija Open Dataset (MIT).
• Entity scheme and training gold derived from DarNERcorp (Moussa & Mourhir, Data in Brief, 2023; CC BY-4.0).
• Silver labelled by an LLM teacher; entity spans kept only where the surface form is verbatim in the source.

Companion models

• AchrafSoltani/camelbert-da-darija-ner (Arabic script) · AchrafSoltani/xlmr-arabizi-darija-ner (Arabizi).

Citation

Paper forthcoming (IANLP / ArabicNLP). Until then, please cite this repository, DarNERcorp, and DODa.