uznlp-uz/uz_medner

UzMedNER

Dataset Summary

UzMedNER is a token-level named entity recognition dataset for Uzbek medical and clinical text. The current release contains 20,133 annotated token rows grouped into 1,310 sentences from 945 documents. Entity annotations use the BIO/IOB2 scheme and cover 31 entity types.

The repository provides both the original Uzbek token data and a row-aligned bilingual variant with an English token-level translation or gloss. It also includes a compact tagset reference file.

Supported Tasks

• medical named entity recognition
• token classification and sequence labeling
• clinical information extraction
• Uzbek medical terminology and normalization research
• cross-lingual or bilingual medical NLP using the English token-gloss column

Files and Configurations

Configuration	File	Rows	Columns	Description
default	UzMedNER.tsv	20,133	7	Primary Uzbek token-level NER data
bilingual	UzMedNER_token_en_filled.tsv	20,133	8	Row-aligned data with Uzbek and English token columns
tagset	tagset.tsv	25	3	Definitions for the 25 core entity types

All three files are UTF-8 encoded and tab-separated. Each configuration currently contains one train split. No predefined validation or test split is included.

Dataset Statistics

Statistic	Value
Token rows	20,133
Documents	945
Sentences	1,310
Entity spans	2,926
Entity-token rows	4,382
Outside (O) rows	15,751
Observed entity types	31
Distinct BIO labels, including O	58
Average sentence length	15.37 tokens
Minimum / maximum sentence length	1 / 129 tokens

An entity span is counted by its B- label. Entity tokens make up approximately 21.77% of all token rows.

Language and Script Distribution

The lang value is assigned at token level.

lang	Rows	Share
uz-Latn	19,480	96.76%
uz-Cyrl	650	3.23%
other	3	0.01%

Dataset Structure

Primary Configuration

The default configuration exposes these fields:

Field	Type	Description
doc_id	string	Document identifier
sent_id	string	Sentence identifier local to a document; combine with doc_id for a unique sentence key
tok_id	integer	One-based token position within the sentence
token	string	Original Uzbek token
tag	string	BIO/IOB2 named entity label, or O for a non-entity token
norm_form	string	Provided normalized or lemma-like form; - means no normalized form is supplied
lang	string	Token-level language/script label: uz-Latn, uz-Cyrl, or other

Bilingual Configuration

The bilingual configuration has the same row order, identifiers, labels, normalization values, and language values as the primary configuration. It replaces token with two columns:

Field	Type	Description
token-uz	string	Uzbek source token; identical to token in the primary file
token-en	string	English token-level translation or gloss

The other fields are doc_id, sent_id, tok_id, tag, norm_form, and lang. Every row in the current bilingual file has a non-empty token-en value.

Tagset Configuration

tagset.tsv contains an ordinal number, the core entity type, and an Uzbek description with examples. The token files are the authoritative source for the complete set of labels observed in this release.

Annotation Scheme

Labels follow valid IOB2 transitions within each sentence:

• B-TYPE begins an entity span.
• I-TYPE continues an entity span of the same type.
• O marks a token outside any entity span.

For example, the first sentence in the release begins as follows:

tok_id	token	tag	norm_form
1	Bemor	B-PATIENT	bemor
2	2-toifa	B-DISEASE	2-type
3	qandli	I-DISEASE	diabet
4	diabet	I-DISEASE	diabet
5	bilan	O	bilan
6	,	O	-
7	insulin	B-DRUG	insulin
8	10	B-DOSAGE	10
9	birlik	I-DOSAGE	birlik
10	i.v	B-ROUTE	intravenoz
11	kuniga	B-FREQ	kuniga
12	2	I-FREQ	2
13	marta	I-FREQ	marta

Sentence boundaries are represented by doc_id and sent_id, not by blank lines. Reconstruct a sentence by grouping on both fields and sorting numerically by tok_id.

Entity Distribution

Entity type	Spans (B-)	Entity tokens (B- + I-)
ALLERGY	8	9
BODY	540	805
CAUSE	26	46
CHEM	15	17
DEVICE	28	39
DISEASE	452	673
DOCTOR	189	297
DOSAGE	38	67
DRUG	260	297
DURATION	12	33
FORM	4	4
FREQ	19	49
LABTEST	23	44
LOC	2	2
LOCATION	13	19
MEASURE	39	67
NEGATION	35	42
ORG	75	120
PATHOGEN	4	7
PATIENT	218	252
PEOPLE	2	2
PHARM	4	4
PROCEDURE	40	58
PROTEIN	2	2
RISK	16	40
ROUTE	11	16
SYMPTOM	497	718
TEMPORAL	174	361
TITLE	7	12
TREATMENT	159	264
UNCERTAINTY	14	16

The 25 core types documented in tagset.tsv are DISEASE, DRUG, DOSAGE, ROUTE, FREQ, DURATION, MEASURE, ALLERGY, NEGATION, LABTEST, SYMPTOM, BODY, UNCERTAINTY, DOCTOR, PATIENT, DEVICE, TREATMENT, FORM, TEMPORAL, PROCEDURE, ORG, RISK, LOCATION, TITLE, and CAUSE.

Six additional types occur in the token annotations: CHEM, LOC, PATHOGEN, PEOPLE, PHARM, and PROTEIN.

Loading the Dataset

Install the Hugging Face datasets library, then load the primary configuration:

from datasets import load_dataset

dataset = load_dataset("uznlp-uz/uz_medner", split="train")
print(dataset[0])

Load the bilingual data or tagset explicitly:

from datasets import load_dataset

bilingual = load_dataset(
    "uznlp-uz/uz_medner",
    "bilingual",
    split="train",
)

tagset = load_dataset(
    "uznlp-uz/uz_medner",
    "tagset",
    split="train",
)

The primary TSV can also be loaded directly:

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files="UzMedNER.tsv",
    delimiter="\t",
    split="train",
)

Reconstructing Sentence-Level Examples

from datasets import load_dataset

rows = load_dataset("uznlp-uz/uz_medner", split="train")
frame = rows.to_pandas().sort_values(["doc_id", "sent_id", "tok_id"])

sentences = (
    frame.groupby(["doc_id", "sent_id"], sort=False)
    .agg(tokens=("token", list), ner_tags=("tag", list))
    .reset_index()
)

When creating evaluation splits, split by doc_id rather than by token row or sentence to reduce document-level leakage.

Intended Use

UzMedNER is intended for research and development in Uzbek medical NLP, including NER model training, sequence-labeling evaluation, terminology extraction, clinical text mining, normalization experiments, and cross-lingual representation learning.

The dataset is not a medical device and must not be used by itself for diagnosis, treatment decisions, or other clinical decision-making.

Data Quality and Validation

For the current release:

• all 20,133 rows have values in every column;
• all (doc_id, sent_id, tok_id) keys are unique;
• tok_id values are consecutive and one-based within every sentence;
• no invalid I- transitions were found under the IOB2 scheme;
• the primary and bilingual files are aligned across all 20,133 rows;
• norm_form uses - as a sentinel on 3,055 rows rather than an empty value.

Limitations

• Entity classes are highly imbalanced: common classes such as BODY, SYMPTOM, and DISEASE have hundreds of spans, while several classes have fewer than ten.
• Most tokens are Uzbek Latin; Uzbek Cyrillic has substantially less coverage.
• Only a train split is released, so published results should document the split strategy and random seed.
• English values are token-level translations or glosses and may not form a fluent sentence when concatenated.
• Uzbek medical language contains spelling variants, abbreviations, borrowed terminology, and productive morphology that may affect generalization.
• The release does not include detailed source-provenance fields, annotator agreement statistics, or a formal de-identification report. Users should review privacy and governance requirements before using the data in sensitive settings.
• tagset.tsv documents 25 core types, while six additional types occur in the token data; users should rely on the observed label inventory above when building models.

Evaluation

For NER experiments, report entity-level precision, recall, and F1 using exact span and entity-type matching. Token-level accuracy may be included as a secondary metric, but it can be dominated by the frequent O label. Keep complete documents together when constructing train, validation, and test partitions.

License

This dataset is released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

Citation

If you use UzMedNER, cite the dataset repository:

@misc{elov_alaev_uzmedner_2026,
  title     = {UzMedNER: An Uzbek Medical Named Entity Recognition Dataset},
  author    = {Elov, Botir B. and Alaev, Ruhillo H.},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/8081},
  url       = {https://huggingface.co/datasets/uznlp-uz/uz_medner},
  license   = {CC BY 4.0}
}