README.md

Spoken-Arabic Named-Entity Dataset (Levantine NER v1.0)

A curated corpus of Levantine-Arabic sentences annotated for Named Entities, plus parallel dual-annotator files for assessing annotation noise.
Ideal for fine-tuning Arabic-BERT–style models on noisy, spoken data and testing cross-register robustness.

---

1 Corpus snapshot

statistic	value
Sentences (unique set)	23 422
Sentences incl. 2nd annotator (A + B)	29 228
Tokens (approx.)	~290 k
Annotated entity spans	17 589
Avg. entities ∕ sentence	0.75
Annotators	Arzy · Rawan · Reem · Sabil · Wiam · Amir
Rounds	round1 – round5 (natural speech) + round6 (synthetic news/MSA)
File format	JSON Lines (UTF-8)

Label inventory

label	description	count
GPE	geopolitical entity	4 601
PER	person	3 628
ORG	organisation	1 426
MISC	misc. named item	1 301
FAC	facility	947
TIMEX	temporal expression	926
DUC	product/brand	711
EVE	event	487
LOC	(non-GPE) location	467
ANG	angle/measure	322
WOA	work of art	292
TTL	title/honorific	227

---

2 File list

file	lines	purpose
unique_sentences.jsonl	23 422	canonical training/dev/test pool (one Levantine sentence per line)
iaa_A.jsonl	5 806	first annotator in each inter-annotator pair (not in unique)
iaa_B.jsonl	5 806	second annotator for the same sentences (aligned 1-to-1 with iaa_A)
sentences.parquet / spans.parquet	52 274 / 17 589	columnar versions for quick Pandas analysis (optional)

Record schema (unique_sentences.jsonl)

{
  "doc_id"    : 137,
  "doc_name"  : "22صدى-الصوت22",
  "sent_id"   : 11,
  "orig_ID"   : "29891",
  "round"     : "round3",            // round1-5 natural, round6 synthetic
  "annotator" : "Rawan",
  "text"      : "جيب جوال أو أي اشي ضو هيك",
  "source_type": "social_videos",
  "spans": [
    { "start": 4, "end": 8, "label": "DUC" }
  ],

  // only for round6
  "msa": {
    "text"  : "<parallel MSA sentence>",
    "spans" : [{ "start": 5, "end": 16, "label": "LOC" }]
  },

  // provenance (optional)
  "url"  : "https://…",
  "date" : "2019-05-02 18:30:44"
}