Persian-NER-Dataset-500k

This repository contains a comprehensive Persian Named Entity Recognition (NER) dataset with approximately 500,000 tokens. This dataset is a collection of all available Persian NER datasets, carefully cleaned and consolidated to ensure the highest quality for training, validating, and testing NER models in the Persian language. The dataset is divided into three subsets: training, validation, and test.

Updating...

Dataset Overview

The dataset is structured as follows:

• Training Set:

  • Number of rows: 446,243
  • Features: ner_tags, tokens, ner_tags_index

• Validation Set:

  • Number of rows: 55,780
  • Features: ner_tags, tokens, ner_tags_index

• Test Set:

  • Number of rows: 55,781
  • Features: ner_tags, tokens, ner_tags_index

Features

• ner_tags: The Named Entity Recognition tags associated with each token.
• tokens: The tokens (words) that have been tagged.
• ner_tags_index: The index values corresponding to the NER tags, which map to specific named entities.

NER Tags

The dataset includes the following NER tags:

• EVENT: Events, including historically significant events or recurring cultural phenomena, etc.
• NORP: Nationalities, religious or political groups, etc.
• TIME: Times less specific than dates, including years, seasons, and general times of day, etc.
• MONEY: Monetary values, including unit, etc.
• DATE: Absolute or relative dates or periods, etc.
• MED: Medicines, including drugs and their categories, etc.
• WORK_OF_ART: Titles of books, songs, and other creative works, etc.
• PERCENT: Percentage expressions, etc.
• FAC: Buildings, airports, highways, bridges, and other facilities, etc.
• LANGUAGE: Any named language, etc.
• ORG: Organizations, including companies, institutions, government bodies, etc.
• PERSON: People, including fictional, etc.
• GPE: Countries, cities, states, etc.
• LAW: Named documents made into laws, etc.
• QUANTITY: Measurements that are not monetary or percentages, etc.
• MISC: Miscellaneous entities, catch-all for entities that do not fit in the above categories, etc.
• DISEASE: Names of diseases and medical conditions, etc.
• PRODUCT: Objects, vehicles, foods, etc., that are products or commodities, etc.
• ORDINAL: "First", "second", etc.
• CARDINAL: Cardinal numbers, etc.

Tag Counts

Below is a table summarizing the counts of each NER tag across the training, validation, and test datasets:

NER Tag	Train	Validation	Test
B-ORG	74,751	9,177	9,397
I-ORG	93,323	11,065	11,581
B-PRODUCT	35,624	4,515	4,441
I-PRODUCT	31,539	3,995	3,975
B-PERSON	160,596	19,785	19,932
I-PERSON	161,614	19,912	20,195
B-NORP	44,956	5,445	5,751
I-NORP	58,288	6,986	7,502
B-GPE	133,887	16,638	16,517
I-GPE	73,628	8,993	8,987
B-WORK_OF_ART	65,521	8,268	8,134
I-WORK_OF_ART	84,308	10,770	10,461
B-CARDINAL	6,560	880	875
I-CARDINAL	2,059	243	258
I-GPE	73,628	8,993	8,987
B-MED	9,347	1,173	1,177
B-EVENT	1,885	233	231
I-EVENT	4,422	579	532
B-FAC	10,167	1,305	1,261
I-FAC	12,465	1,599	1,501
B-DATE	11,712	1,407	1,451
I-DATE	12,323	1,537	1,544
B-QUANTITY	632	79	78
I-QUANTITY	1,162	155	146
B-MONEY	2,514	325	331
I-MONEY	5,383	738	726
B-TIME	1,468	179	200
I-TIME	1,844	234	266
B-DISEASE	8,524	1,072	1,021
B-ORDINAL	1,486	181	182
I-DISEASE	8,783	1,135	1,086
B-LAW	241	46	50
I-MED	4,457	605	600
B-MISC	3,477	474	435
I-LAW	498	103	131
B-PERCENT	1,895	249	190
I-PERCENT	2,340	313	243
I-MISC	1,819	238	210
B-LANGUAGE	280	38	25
I-ORDINAL	19	6	3
I-LANGUAGE	19	4	6

Usage

This dataset can be used to train, validate, and test NER models in the Persian language. The structured format of the dataset allows easy integration into various deep learning frameworks like TensorFlow, PyTorch, or Hugging Face's transformers.

Example Code

Here's an example of how you might load and use this dataset in Python using the datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("mansoorhamidzadeh/Persian-NER-Dataset-500k")

# Access the train, validation, and test sets
train_dataset = dataset['train']
validation_dataset = dataset['validation']
test_dataset = dataset['test']

# Example: Accessing the first example from the train set
print(train_dataset[0])

@dataset{hamidzadeh2024persian_ner_500k,
  author       = {Mansoor Hamidzadeh},
  title        = {Persian-NER-Dataset-500k},
  year         = 2024,
  url          = {https://huggingface.co/datasets/mansoorhamidzadeh/Persian-NER-Dataset-500k}
}