ShamNER – Spoken Arabic Named‑Entity Recognition Corpus (Levantine v1.1)

ShamNER is a curated corpus of Levantine‑Arabic sentences annotated for Named Entities, plus dual annotation to check for consisetency (agreement) across human annotators.

• Rounds : pilot, round1–round5 (manual, as a rule quality improved across rounds) and round6 (synthetic, post‑edited). The sythentic data is done by sampling label-rich annotated spans from an MSA project and writing it with an LLM while force-injecting the annotated spans. Native speakers of Arabic then edited the these chunks to see to it that they sound as fluent and dilactical as possible. They were instructed not to touch the annotated spans. A script validated that no spans were modified.
• Strict span‑novel evaluation : validation and test contain no entity surface‑form that appears in train (after normalisation). This probes true generalisation.
• Tokeniser‑agnostic : only raw sentences and character spans are stored; regenerate BIO tags with any tokenizer you wish.

Quick start

# Uncomment next line if you hit a LocalFileSystem / fsspec error on Colab
# !pip install -U "datasets>=2.16.0" "fsspec>=2023.10.0"
from datasets import load_dataset
sham = load_dataset("HebArabNlpProject/ShamNER")
train_ds = sham["train"]

datasets streams the top‑level *.parquet files automatically; use the matching *.jsonl for grep‑friendly inspection.

Split Philosophy

• No duplicate documents – A document is identified by the pair
  (doc_name, round); each such bundle is assigned to exactly one split. This rule holds true for bundles, though individual sentences within bundles might have overlapping spans after post-allocation pruning for specific thresholds.

• Rounds – Six annotation iterations:
  pilot, round1 – round5 (manual, quality improving each round) and
  round6 (synthetic, then post-edited).
  Early rounds feed train; span-novel slices of round5 + round6 populate test.

• Single test set – The corpus ships one held-out test split:
  test = span-novel bundles from round 5 plus span-novel bundles from round 6.
  No separate test_synth file.

• Span-novelty rule (Relaxed)

Before allocation, normalise every entity string:

• Convert to lowercase (Latin aphbaet exists in social media)
• Strip Arabic diacritics
• Remove leading “ال”
• Collapse internal whitespace

A bundle is forced to train if any of its normalised spans already occurs in train.

A post-allocation pruning step then moves sentences from validation or test back to train
only if more than 50% of their normalized spans already exist in the training set.

This threshold (0.50) was chosen to provide more learning examples to the model
in the evaluation sets, leading to improved performance.

• Tokeniser-agnostic – Each record stores only raw text and character-offset spans; no BIO arrays. Users regenerate token-level labels with whichever tokenizer their model requires.

Split sizes

split	sentences	files
train	19 532	train.jsonl / train.parquet
validation	1 931	validation.*
test	1 931	test.*
iaa_A	5 806	optional, dual annotator A
iaa_B	5 806	optional, annotator B

Every sentence that appears in iaa_A.jsonl is also in the train split (with the same labels), while iaa_B.jsonl provides the alternative annotation for agreement/noise studies.

Label inventory (computed from unique_sentences.jsonl)

label	description	count
GPE	Geopolitical Entity	4 601
PER	Person	3 628
ORG	Organisation	1 426
MISC	Catch-all category	1 301
FAC	Facility	947
TIMEX	Temporal expression	926
DUC	Product / Brand	711
EVE	Event	487
LOC	(non-GPE/natural) Location	467
ANG	Language	322
WOA	Work of Art	292
TTL	Title / Honorific	227

File schema (*.jsonl)

{
  "doc_id": 137,
  "doc_name": "mohamedghalie",
  "sent_id": 11,
  "orig_ID": 20653,
  "round": "round3",
  "annotator": "Rawan",
  "text": "جيب جوال أو أي اشي ضو هيك",
  "spans": [
    {"start": 4, "end": 8, "label": "DUC"}
  ]
}

Inter‑annotator files

iaa_A.jsonl and iaa_B.jsonl contain parallel annotations for the same 5 806 sentences. Use them to measure agreement or experiment with noise‑robust training. These sentences do not overlap with the primary train/val/test splits. As stated above, only iaa_A.jsonl were injected into the train, dev and test set.

© 2025 · CC BY‑4.0