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)
• Format: JSONL / Parquet
• Records: 3,700,048
• Total Tokens: 47,529,930
• Entity Tags:
  • B-PER, I-PER – Person
  • B-ORG, I-ORG – Organization
  • B-LOC, I-LOC – Location
  • O – Outside any entity

Tag Distribution

Tag	Count
O	44,442,959
B-PER	1,376,902
I-PER	234,671
B-ORG	1,258,079
I-ORG	35,966
B-LOC	177,360
I-LOC	3,993

Sentence Statistics

Metric	Count
Total Sentences	3,700,048
Sentences with Tags	1,867,708
Sentences with only O	1,832,340
Avg Tokens per Sentence	12.85

Construction Steps

1. Extracted all tokens from raw Dhivehi sentence dataset.
2. Cleaned while retaining Thaana and Arabic character ranges.
3. Ranked tokens by frequency and built a vocabulary.
4. Matched tokens to entity dictionaries (PER, ORG, LOC) from CSVs.
5. Assigned raw tags (PER, ORG, LOC) to matched tokens.
6. Converted to BIO format with proper B-/I- tagging.
7. Exported final JSONL and Parquet datasets for use with Hugging Face.

Accuracy Notice

This dataset was automatically generated using a weakly-supervised method and large lookup lists. It was not manually reviewed or annotated. Some entity boundaries and false positives/negatives are expected.

Use for:

• Pretraining NER models
• Weak supervision pipelines
• Dhivehi-specific NLP research

Dataset Preview

Each sample has:

{
  "id": 123,
  "sentence": "މުޙައްމަދު އަރޝާދު ކުރިއަށް",
  "tokens": ["މުޙައްމަދު", "އަރޝާދު", "ކުރިއަށް"],
  "ner_tags": [1, 2, 0]
}