uznlp-uz/uzbek_NER

Uzbek NER Gold

Uzbek NER Gold is a token-level named entity recognition dataset for Uzbek. The dataset is distributed as a UTF-8 TSV file and uses BIO tagging for named entities.

Dataset Summary

• Dataset ID: uznlp-uz/uzbek_NER
• Language: Uzbek (uz)
• Rows: 59,569 token rows
• Columns: 5
• Sentences: 4,176
• Split: train
• Format: UTF-8 TSV
• Data file: Uzbek_NER_Gold.tsv
• License: CC BY 4.0

Data Fields

Field	Description
Sentence	Sentence identifier. Tokens with the same value belong to the same sentence.
TokenOrder	1-based token position inside the sentence.
Token	Token text.
NER_Tag	BIO named entity tag. O marks tokens outside named entities.
pos	Source POS or token category label.

Tagset

The dataset uses BIO labels over the following named entity types.

Entity	Tag	Gold v1.0 guideline
Person	PER	Inson ismi yoki aniq shaxs nomi; lavozim, kasb, umumiy guruh va olmoshlar PER emas.
Organization	ORG	Tashkilot, vazirlik, universitet, agentlik, kompaniya, qo‘mita, fond/jamg‘arma va boshqalar.
Location	LOC	Davlat, respublika, viloyat, shahar, tuman, geografik joy nomlari.
Miscellaneous	MISC	Qolgan maxsus nomlangan obyektlar; aniq sinfga tushmaydigan nomlar.
Money	MONEY	Pul birliklari va pul miqdorlari.
Number	NUMERIC	Son, raqam, miqdor ifodalari, money/time bo‘lmagan raqamli birliklar.
Date/Time	TEMPORAL	Sana, vaqt, davr, yil, oy, kun va vaqt oralig‘i.
Work	WORK	Asar, kitob, film, loyiha, badiiy yoki ilmiy ish nomlari; standart yozilishi WORK.

Label Values

NER labels: O, B-PER, I-PER, B-ORG, I-ORG, B-LOC, I-LOC, B-MISC, I-MISC, B-MONEY, I-MONEY, B-NUMERIC, I-NUMERIC, B-TEMPORAL, I-TEMPORAL, B-WORK, I-WORK

POS/category labels: C, IB, II, JJ, MD, N, NER, NUM, P, PUNCT, Prt, RR, UH, UNK, VB

Statistics

Overview

Metric	Value
Token rows	59,569
Columns	5
Sentences	4,176
Unique tokens, case-sensitive	14,791
Unique tokens, case-folded	13,969
Duplicate rows	0
Empty values	0
Minimum sentence length	1
Maximum sentence length	253
Mean sentence length	14.26
Median sentence length	12

Entity Span Distribution

Entity type	Gold v1.0 spans	Original spans	Difference
LOC	2,302	2,262	40
ORG	2,148	2,207	-59
PER	1,653	2,845	-1,192
MISC	1,400	1,498	-98
TEMPORAL	591	615	-24
NUMERIC	581	591	-10
WORK	441	465	-24
MONEY	134	140	-6
Total	9,250	10,623	-1,373

NER Tag Distribution

NER tag	Count
O	43,445
B-LOC	2,302
I-ORG	2,191
B-ORG	2,148
B-PER	1,653
B-MISC	1,400
I-MISC	1,258
I-PER	1,161
I-LOC	842
B-TEMPORAL	591
B-NUMERIC	581
I-WORK	448
B-WORK	441
I-NUMERIC	437
I-TEMPORAL	323
I-MONEY	214
B-MONEY	134

POS/Category Distribution

POS/category	Count
N	17,733
NER	12,914
VB	9,746
PUNCT	4,628
JJ	4,096
IB	2,676
II	2,137
C	1,478
P	1,328
RR	1,207
NUM	956
MD	417
Prt	225
UNK	26
UH	2

Normalization

• Text is stored as UTF-8.
• Uzbek text is normalized to Latin script.
• Uzbek apostrophes were normalized to distinct Unicode characters: ‘ for o‘/g‘ and ’ for the tutuq sign.
• Non-standard apostrophe variants were removed from the released TSV.
• Each token is stored on a separate TSV row.
• No field contains empty values in the released file.

Loading

from datasets import load_dataset

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

The TSV file can also be loaded directly:

from datasets import load_dataset

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

Intended Use

This dataset can be used for Uzbek named entity recognition research and development, including token classification, NER model training, NER evaluation, information extraction, and benchmarking Uzbek language models.

Notes

The dataset is sentence-aware through the Sentence and TokenOrder fields. To reconstruct sentences, group rows by Sentence and sort them by TokenOrder.

The NER_Tag field follows BIO notation. A tag beginning with B- starts an entity span, and a tag beginning with I- continues an entity span of the same type.

Citation

If you use Uzbek NER Gold, cite the dataset repository:

@misc{uzbek_ner_gold,
  title = {Uzbek NER Gold},
  author = {{Elov B.B., Alaev R.H.}},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/uznlp-uz/uzbek_NER}},
  license = {CC BY 4.0}
}