README.md · iahlt/arabic_ner

metadata

task_categories:
  - token-classification
language:
  - ar
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: tokens
      sequence: string
    - name: raw_tags
      sequence: string
    - name: ner_tags
      sequence: int64
    - name: spaces
      sequence: int64
    - name: spans
      list:
        - name: end
          dtype: int64
        - name: label
          dtype: string
        - name: start
          dtype: int64
        - name: text
          dtype: string
    - name: record
      dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 181231147
      num_examples: 40000
  download_size: 52900580
  dataset_size: 181231147

Arabic NER

This dataset provides spans-level and token-level Named Entity Recognition (NER) annotations for Arabic text.

It includes:

Coarse-grained annotations (e.g., MISC, ORG, LOC, etc.) appearing in the spans field.
Fine-grained annotations in BILUO format (e.g., B-MISC, I-MISC, L-MISC) in the raw_tags field, and their corresponding numeric IDs in the ner_tags field.

Check annotation guideline and details here

Below is a brief description of each field:

tokens: A list of individual tokens (words) extracted from the text. Important note, it's processed using span-based annotation(spans), to get morhpology-rich token-based annotation, please use your own tokenization and processing method. 
raw_tags: The original, text-based BILUO entity tags per token (e.g., B-MISC, I-MISC, etc.).
ner_tags: The integer indices corresponding to each raw_tags label (following the label scheme in dataset_info).
spaces: A sequence of integers (0 or 1) indicating whether a space follows the corresponding token.
spans: A list of dictionaries, each containing (top-level of the annotated record["label_hierarchy"] which contains both annotations fine- and coarse-grained):
    start and end character offsets of the entity in the original text
    text of the entity span
    label (entity)
record: A JSON-encoded string with additional metadata about the document ['metadata', 'text', 'label', 'user', 'timestamp', 'flatten', 'label_hierarchy', 'has_overlappings', 'n_hierarchy_levels']

Sample of the Arabic NER data:

Usage

pip install datasets

huggingface-cli login

from datasets import load_dataset


ds = load_dataset("iahlt/arabic_ner_mafat")

for sample in ds["train"]:
   print(sample)

Sample

{
 "tokens": ["يجب", "عليك", "الامتناع", "عن", "مضغ", "العلكة", "ا", "ٕ", "ذا", "كنت", "تعاني", "من", "ا", "ٔ", "ي", "نوع", "من", "الام", "الفك", "ا", "ٔ", "و", "اضطراب", "الصدغي", "الفكي", "."],
 "raw_tags": ["O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "B-MISC", "I-MISC", "L-MISC", "O"],
 "ner_tags":  [32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 30, 28, 31, 32],
 "spaces": [1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0],
 "spans": [
  {
   "end": 94,
   "label": "MISC",
   "start": 75,
   "text": "اضطراب الصدغي الفكي"
  }
 ],
 "record": "{\"metadata\": {\"doc_id\": \"0142895c6cdb030b10c8cc2e5c9639f9422bf22ef45a1b314d7a366fc6489938\", \"url\": \"https://www.alarab.com//Article/1004953\", \"source\": \"AlArab\", \"title\": \"فوائد غير متوقعة للعلكة الخالية من السكر.. اكتشفوها معنا!\", \"authors\": \"كل العرب (تصوير: iStockphoto)\", \"date\": \"2021-08-30 13:25:01\", \"domains\": \"التغذية الصحيحة:فوائد العلكة الخالية من السكر\", \"parnumber\": \"36\", \"sentnumber\": \"1\", \"manually_qa-ed\": \"Yes\"}, \"text\": \"يجب عليك الامتناع عن مضغ العلكة إذا كنت تعاني من أي نوع من الام الفك أو اضطراب الصدغي الفكي.\", \"label\": [[75, 94, \"MISC\"]], \"user\": \"nlhowell\", \"timestamp\": 1685356359.342268, \"flatten\": {\"tokens\": [\"يجب\", \"عليك\", \"الامتناع\", \"عن\", \"مضغ\", \"العلكة\", \"ا\", \"ٕ\", \"ذا\", \"كنت\", \"تعاني\", \"من\", \"ا\", \"ٔ\", \"ي\", \"نوع\", \"من\", \"الام\", \"الفك\", \"ا\", \"ٔ\", \"و\", \"اضطراب\", \"الصدغي\", \"الفكي\", \".\"], \"ner_tags\": [\"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"O\", \"B-MISC\", \"I-MISC\", \"L-MISC\", \"O\"], \"spaces\": [1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0]}, \"label_hierarchy\": {\"0\": [{\"end\": 94, \"label\": \"MISC\", \"start\": 75, \"text\": \"اضطراب الصدغي الفكي\"}], \"1\": null, \"2\": null, \"3\": null, \"4\": null, \"5\": null}, \"has_overlappings\": false, \"n_hierarchy_levels\": 1}"
}

Visualization

pip install spacy ipython -q

import json
from spacy.training import offsets_to_biluo_tags, biluo_tags_to_spans

record = ds[676]
record["record"] = json.loads(record["record"])

ner_tags = record["raw_tags"]
tokens = record["tokens"]


doc = spacy.tokens.Doc(spacy.blank("ar").vocab, words=tokens)
doc.ents = biluo_tags_to_spans(doc, ner_tags)

print(record["record"]["text"])
spacy.displacy.render(doc, style="ent", jupyter=True)