chuuhtetnaing/myanmar-ner-dataset

Please visit the GitHub repository for other Myanmar Language datasets.

Myanmar NER Dataset

A token classification dataset for Myanmar (Burmese) Named Entity Recognition (NER), formatted for sequence labeling tasks using BIO tagging scheme.

📓 Data Preparation Notebook: dataset-preparation.ipynb

📓 Fine-Tuning Notebook: myanmar-ner-fine-tuning.ipynb (based on the HuggingFace Token Classification Guide)

Dataset Description

This dataset contains Myanmar language text annotated with Named Entity Recognition (NER) tags. The dataset has been cleaned, preprocessed, and formatted to be ready for fine-tuning HuggingFace token classification models.

The dataset uses a 7-tag annotation scheme (6 entity types + O tag) designed to support various NLP tasks including information extraction, question answering, and knowledge graph construction.

Note: The original myNER corpus uses the BIOES (Begin, Inside, Outside, End, Single) tagging scheme. This preprocessed version has been converted to the BIO (Begin, Inside, Outside) format for easier fine-tuning with HuggingFace Token Classification models.

Source Data

Original Dataset: ye-kyaw-thu/myNER

Original Corpus: The myNER corpus containing 16,605 sentences was originally sourced from the myPOS version 3 corpus and manually annotated for NER.

This is a preprocessed and cleaned version of the original myNER dataset. Please cite the original dataset and authors if you use this preprocessed version in your research.

Processing Pipeline

1. Data Loading: Loaded CoNLL format data from the original myNER-7tags corpus
2. Syllable Segmentation: Applied Myanmar syllable breaking for proper tokenization
3. Tag Conversion & Expansion: Converted BIOES to BIO and expanded tags for multi-syllable words (B/S → B+I..., I/E → I..., O → O...)
4. Token-Label Alignment: Verified proper alignment between syllables and their NER tags
5. Deduplication: Removal of duplicate entries
6. Train-Test Split: 80% training (12,825 examples) and 20% test (3,207 examples)
7. Format Conversion: Conversion to HuggingFace token classification format

Tagging Scheme Conversion

The original corpus uses word-level BIOES tagging. After syllable segmentation, each word may become multiple syllables, so tags are expanded accordingly:

BIOES (Original)	BIO (After Syllable Segmentation)	Description
B-	B-, I-, I-, ...	Begin → B for first syllable, I for remaining syllables
I-	I-, I-, I-, ...	Inside → all syllables remain I
O	O, O, O, ...	Outside → all syllables remain O
E-	I-, I-, I-, ...	End → all syllables become I
S-	B-, I-, I-, ...	Single → B for first syllable, I for remaining syllables

Dataset Statistics

Split	Examples
Train	12,825
Test	3,207

Data Format

{
    "tokens": [ "ရွာ", "နေ", "ရာ", "ကုတ်", "မှာ", "၁", "၇", "၆", "၉", "၂", "၄", "ဖြစ်", "သည်", "။" ],
    "ner_tags": [ 12, 12, 12, 12, 12, 4, 5, 5, 5, 5, 5, 12, 12, 12 ]
}

Features

• tokens: Sequence[string] - Input tokens (Myanmar syllables)
• ner_tags: Sequence[ClassLabel] - Named Entity Recognition labels

NER Tags

The dataset includes 6 entity types using BIO tagging scheme:

Entity Type	Description
DATE	Date expressions
LOC	Location
NUM	Numerical expressions
ORG	Organization
PER	Person
TIME	Time expressions

Tag Labels

The full tag list in BIO format:

Tag	Index
B-DATE	0
I-DATE	1
B-LOC	2
I-LOC	3
B-NUM	4
I-NUM	5
B-ORG	6
I-ORG	7
B-PER	8
I-PER	9
B-TIME	10
I-TIME	11
O	12

Usage

from datasets import load_dataset
ds = load_dataset("chuuhtetnaing/myanmar-ner-dataset")

ds["train"].features["ner_tags"].feature.names
# ['B-DATE', 'I-DATE', 'B-LOC', 'I-LOC', 'B-NUM', 'I-NUM', 'B-ORG', 'I-ORG', 'B-PER', 'I-PER', 'B-TIME', 'I-TIME', 'O']

Example from the dataset:

ds["train"][0]
# {
#     'tokens': [ "1", "bit", "ပုံ", "ရိပ်", "တစ်", "ခု", "သည်", "monochrome", "ဖြစ်", "သည်", "။" ],
#     'ner_tags': [ 4, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12 ]
# }

id2label = {i: label for i, label in enumerate(ds["train"].features["ner_tags"].feature.names)}
label_ner_tags = [id2label[tag_id] for tag_id in ds["train"][0]["ner_tags"]]

Intended Use

• Training NER models for Myanmar NLP
• Token classification / sequence labeling experiments (HuggingFace Token Classification Training Example)
• Myanmar language processing research
• Information extraction and knowledge graph construction
• Question answering systems
• Machine translation preprocessing

Citation

arXiv Preprint: arXiv:2504.04038

For more information about the original dataset, please visit:

• GitHub Repository
• arXiv Paper

Acknowledgments

Special thanks to:

• Professor Ye Kyaw Thu and the Language Understanding Lab (LU Lab) team for creating the original myNER dataset
• The Myanmar NLP community for supporting language technology development