myFineNER: A Fine-Grained Burmese NER Benchmark

myFineNER is a fine-grained named entity recognition (NER) corpus for Burmese. It contains 4,918 sentences from two domains, formal news-style text and spoken/colloquial text. Every token is annotated with a part-of-speech (POS) tag and a named entity tag in the BIOES scheme, using 18 entity types.

• Code: https://github.com/cpe-nlp-lab/myFineNER
• Paper: myFineNER: A Fine-Grained Burmese NER Benchmark with Transformer Models and Multilingual LLM Evaluation (2026, under review)
• Curated by: H. M. Thu, P. P. Thit, P. B. Moe and A. H. Khine
• Language: Burmese (my)
• License: CC BY-NC-SA 4.0

Dataset Summary

Domain	Sentences	Tokens	Entity mentions
Formal	2,563	95,914	12,048
Spoken	2,355	35,094	5,293
Combined	4,918	131,008	17,341

Formal sentences are considerably longer than spoken ones (about 37 tokens per sentence against about 15), so the formal domain accounts for most of the tokens although the two domains have a similar number of sentences.

Uses

Direct Use

• Training and evaluating Burmese NER models, including fine-grained NER.
• Training and evaluating Burmese POS taggers, or joint POS and NER models.
• Studying domain transfer between formal and spoken Burmese.
• Evaluating multilingual large language models on Burmese sequence labelling.

Out-of-Scope Use

• Commercial use of any kind. The license permits non-commercial use only.
• Identifying, profiling or monitoring the individuals named in the text.

Dataset Structure

Files

combined/  train.conll  dev.conll  test.conll
formal/    burmese_ner_pos_Official.conll  formal_train.conll  formal_dev.conll  formal_test.conll
spoken/    burmese_ner_pos_Spoken.conll    spoken_train.conll  spoken_dev.conll  spoken_test.conll

• formal/burmese_ner_pos_Official.conll and spoken/burmese_ner_pos_Spoken.conll are the full corpora for each domain.
• formal/ and spoken/ also hold the train, dev and test split of each domain.
• combined/ holds the union of the two domains for each split, shuffled. These are the splits used for the main results in the paper.

Format

Files are in CoNLL format. Each line holds one token with three tab-separated columns: the token, its POS tag and its NER tag. A blank line separates sentences.

မိုခါမုန်တိုင်း	n	S-DISASTER
ဝင်	v	O
မယ်	part	O
ဆို	v	O
လို့	part	O
ကြိုတင်ပြင်ဆင်မှု	n	O
တွေ	part	O
လုပ်	v	O
နေ	part	O
ကြ	part	O
ပြီ	part	O

Burmese is written without spaces between words, so the corpus is distributed already segmented into words. Models should use this tokenization as given.

Splits

Each domain is split 70/15/15 into train, dev and test, stratified by entity type with a fixed seed (42).

Split	Sentences	Tokens	Entity mentions
Formal train	1,794	69,711	9,144
Formal dev	384	13,183	1,445
Formal test	385	13,020	1,459
Spoken train	1,648	26,240	4,225
Spoken dev	353	4,440	536
Spoken test	354	4,414	532
Combined train	3,442	95,951	13,369
Combined dev	737	17,623	1,981
Combined test	739	17,434	1,991

Named Entity Tags

Tags follow the BIOES scheme: B- begins a multi-token entity, I- is inside it, E- ends it, S- marks a single-token entity, and O marks a token outside any entity. With 18 entity types this gives 73 labels.

Type	Description
PER	Personal names, without honorifics or titles
TITLE	Honorifics and academic, religious, military or professional titles attached to a specific name
JOB	Professions and work roles that are not a formal title attached to a name
GPE	Administrative places: countries, states, regions, cities, towns, villages
LOC	Non-administrative geographic locations: rivers, mountains, seas, forests, natural regions
FAC	Man-made facilities and structures: pagodas, bridges, airports, hospitals, buildings
ORG	Organizations: companies, ministries, departments, schools, political parties, NGOs, armed groups
DATE	Dates, years, months, days and relative date expressions
TIME	Clock times and durations
MONEY	Monetary amounts, including the number and the currency unit
ETHNIC	Ethnic groups, nationalities and cultural identity groups
REL	Religions, religious sects and belief systems
DISEASE	Diseases, viruses, syndromes and health conditions
DISASTER	Natural disasters and disaster types
PRODUCT	Brands, models, software, devices and specific manufactured products
FOOD	Dishes, drinks and specific food items
EVENT	Named historical, political or social events, excluding festivals
FEST	Recurring festivals and celebrations

Grammatical particles, postpositions and plural markers are tagged O unless they are part of an entity's name.

Part-of-Speech Tags

Tag	Meaning	Tag	Meaning
abb	Abbreviation	part	Particle
adj	Adjective	ppm	Postpositional marker
adv	Adverb	pron	Pronoun
conj	Conjunction	punc	Punctuation
fw	Foreign word	sb	Symbol
int	Interjection	tn	Text number
n	Noun	v	Verb
num	Number

How to Use

Download the files:

huggingface-cli download cpe-kmutt-nlp/myFineNER --repo-type dataset --local-dir data

Access is gated. Accept the terms on this page and run huggingface-cli login first if the download is refused.

The scripts for every experiment in the paper are in the code repository.

Dataset Creation

Curation Rationale

Burmese is a low-resource language for NER: annotated data is scarce, and fine-grained entity types and colloquial text are especially under-represented. myFineNER was created to provide a fine-grained benchmark that covers both formal and spoken Burmese.

Source Data

• Formal subset. Derived from publicly accessible Burmese news outlets.
• Spoken subset. Created by consenting native Burmese speakers.

Annotation Process

Annotation combined LLM drafting with human correction:

1. Word segmentation. Sentences were segmented into words with Gemini-2.5-Flash under a detailed segmentation prompt followed by the human review.
2. Draft annotation. Gemini-2.5-Flash produced draft POS tags and draft BIOES entity tags for the fixed token sequence, following written tagging guidelines.
3. Human correction. Human annotators reviewed and corrected the drafts. The released labels are the corrected ones.
4. Quality assessment. Three annotators independently labelled a 500-sentence sample, and inter-annotator agreement was measured with Fleiss' and Cohen's kappa. The reliability of the Gemini drafts was measured against the human-corrected labels.

The prompts containing the full segmentation and tagging guidelines are in the code repository.

Who are the annotators?

Annotation was performed by the authors and native-speaker collaborators. No crowdsourced labor was used.

Personal and Sensitive Information

The corpus contains names of people, organizations and places, which are the target of the NER annotation. No personally identifiable information beyond what appears in publicly reported news is included.

Bias, Risks, and Limitations

• Domain coverage. The corpus covers formal news-style and spoken text. Models trained on it may perform worse on other kinds of text, such as social media, legal or medical writing.
• Class imbalance. Entity types are unevenly distributed, and some types have few examples, so per-type scores for rare types are less reliable.
• LLM-drafted labels. Labels were drafted by an LLM before human correction. Systematic errors in the drafts that annotators did not catch may remain.
• Fixed tokenization. Entity boundaries depend on the released word segmentation. A different segmenter will produce tokens that do not align with the labels.

License

The dataset is released under CC BY-NC-SA 4.0. It may be used, shared and adapted for non-commercial purposes, with attribution, and adaptations must be shared under the same license.

Citation

H. M. Thu, P. P. Thit, P. B. Moe, and A. H. Khine, "myFineNER: A fine-grained Burmese NER benchmark with transformer models and multilingual LLM evaluation," 2026. PACLIC.

@inproceedings{thu2026myfinener,
  title  = {{myFineNER}: A Fine-Grained {Burmese} {NER} Benchmark with Transformer Models and Multilingual {LLM} Evaluation},
  author = {Thu, H. M. and Thit, P. P. and Moe, P. B. and Khine, A. H.},
  year   = {2026},
  note   = {40th Pacific Asia Conference on Language, Information and Computation}
}

Dataset Card Contact

Authors