alakxender/dhivehi-ner-dataset

Dhivehi NER Dataset

This dataset is a weakly-supervised Named Entity Recognition (NER) dataset for the Dhivehi language, built from a large unlabeled sentence corpus using dictionary-based tagging and BIO post-processing.

Dataset Summary

• Language: Dhivehi (ދިވެހި) + Arabic (For Dhivehi names only)
• Records: 90,735 (cleaned from 97,308 original)
• Total Tokens: 775,136
• Total Entities: 128,764
• Data Quality: 93.2% (6,573 invalid records removed)
• Average Sentence Length: 8.5 tokens
• Average Entity Length: 1.1 tokens

Entity Types and Labels

The dataset uses the BIO (Beginning-Inside-Outside) tagging scheme:

Label ID	Label Name	Description	Count	Percentage
0	O	Outside any entity	646,372	83.4%
1	B-PER	Beginning of Person	43,973	5.7%
2	I-PER	Inside Person	7,228	0.9%
3	B-ORG	Beginning of Organization	28,401	3.7%
4	I-ORG	Inside Organization	4,910	0.6%
5	B-LOC	Beginning of Location	39,495	5.1%
6	I-LOC	Inside Location	2,211	0.3%
7	B-MISC	Beginning of Miscellaneous	1,861	0.2%
8	I-MISC	Inside Miscellaneous	685	0.1%

Dataset Fields

Each record contains the following fields:

Field	Type	Description
text	string	Original sentence text
token	list[string]	Tokenized words/tokens
ner_tags	list[int]	Numeric entity labels (0-8)
ner_class	list[string]	String entity labels (B-PER, I-ORG, etc.)

Data Quality

The dataset has been cleaned and validated:

• Length Consistency: All records have matching token, tag, and class lengths
• Label Validation: All tags are valid integers (0-8)
• BIO Consistency: Proper B-/I- prefix usage for all entity types
• Invalid Records Removed: 6,573 records with quality issues were excluded

Dataset Preview

{
  "text": "މާދަމާގެ އެއްވުމަށް ފުލުހުން ޝަރުތުތަކެއް ކަނޑައަޅައި އަދާލަތު ޕާޓީ އަށް ސިޓީ ފޮނުވައިފި",
  "token": ["މާދަމާގެ", "އެއްވުމަށް", "ފުލުހުން", "ޝަރުތުތަކެއް", "ކަނޑައަޅައި", "އަދާލަތު", "ޕާޓީ", "އަށް", "ސިޓީ", "ފޮނުވައިފި"],
  "ner_tags": [0, 0, 3, 0, 0, 3, 4, 0, 0, 0],
  "ner_class": ["O", "O", "B-ORG", "O", "O", "B-ORG", "I-ORG", "O", "O", "O"]
}

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("alakxender/dhivehi-ner-dataset")

# Access records
for record in dataset["train"]:
    print(f"Text: {record['text']}")
    print(f"Tokens: {record['token']}")
    print(f"NER Tags: {record['ner_tags']}")
    print(f"NER Classes: {record['ner_class']}")

Accuracy Notice

This dataset was processed using automated cleaning and validation techniques. While quality issues have been addressed, some entity boundaries and classifications may still require manual review for production use.

Recommended Use Cases:

• Pretraining NER models for Dhivehi
• Research and development
• Baseline model training
• Weak supervision pipelines