makhmudlp/uzbek_text_classification_dataset

Uzbek News Sentences Dataset

Overview

A curated sentence-level dataset of Uzbek-language text scraped from two major Uzbek news portals — kun.uz and daryo.uz — and labeled with six topic categories. The dataset is designed for text classification, language modeling, and general NLP research on Uzbek, a Turkic language spoken by ~45 million people.

Why this dataset exists

Uzbek is critically underrepresented in modern NLP and AI training corpora. Most multilingual benchmarks and pre-trained models treat Uzbek as an afterthought, if they include it at all. Consequences include:

• Poor machine translation quality (Uzbek ↔ English / Russian)
• No reliable Uzbek-specific sentence classifiers
• Tokenizers that fragment Uzbek words incorrectly
• Zero-shot performance gaps compared to high-resource languages

This dataset is a first step toward closing that gap by providing clean, deduplicated, topic-labeled Uzbek sentences in Latin script (the official script since 1995), ready to use with the HuggingFace datasets library.

Dataset statistics

Split	Rows
train	517
validation	67
test	59
Total	643

Label distribution (stratified across all splits)

Label	Count	% of total
politics	139	21.6 %
science	115	17.9 %
culture	111	17.3 %
sports	107	16.6 %
education	92	14.3 %
general	79	12.3 %

Average sentence length: 17.6 words (range: 5 – 84 words)

Columns

Column	Type	Description
sentence	string	A single Uzbek sentence in Latin script
label	ClassLabel	Topic category (integer index; see class names below)
source	string	Origin website — "kun.uz" or "daryo.uz"
language	string	ISO 639-1 language code — always "uz"

Class label mapping

0 → culture
1 → education
2 → general
3 → politics
4 → science
5 → sports

Data collection & cleaning

1. Scraping — articles were fetched from kun.uz category feeds (HTML) and the daryo.uz REST API (data.daryo.uz/api/v1/site/news/by-category).
2. Extraction — article body text selected via CSS selector div.news-inner__content-page (kun.uz) and short_content API field (daryo.uz); navigation, ads, and share widgets removed before parsing.
3. Sentence splitting — split on terminal punctuation (., !, ?, …) followed by a capital letter or Uzbek character.
4. Quality filters applied per sentence:
   • Minimum 5 words and 25 characters
   • Maximum 84 words and 600 characters
   • ≤ 30 % non-alphabetic characters
   • ≤ 25 % digit characters
   • Must contain Uzbek-specific letter combinations (oʻ, gʻ, sh, ch, ng) or ≥ 2 common Uzbek function words
5. Deduplication — exact-match dedup on lowercased, whitespace-normalized sentence strings.
6. Script standardisation — Uzbek Cyrillic detected and transliterated to Latin using the official 1995 Uzbek Latin alphabet standard (no Cyrillic sentences were present in this collection, but the pipeline handles them).
7. Labeling — two-stage hybrid:
   • If ≥ 2 keyword hits for a specific category → use that category
   • Otherwise fall back to the source feed's category as a prior (sport feed → sports, technology feed → science, etc.)

How to load

From local disk

from datasets import load_from_disk

ds = load_from_disk("./uzbek_sentences_dataset")
print(ds)
# DatasetDict({
#     train:      Dataset(643 rows, features: [sentence, label, source, language])
#     validation: Dataset(...)
#     test:       Dataset(...)
# })

# Decode integer label back to string
label_names = ds["train"].features["label"].names  # ['culture', 'education', ...]
example = ds["train"][0]
print(example["sentence"])
print(label_names[example["label"]])

From HuggingFace Hub (after upload)

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/uzbek-sentences")

Quick classifier fine-tune with transformers

from datasets import load_from_disk
from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
import numpy as np

ds         = load_from_disk("./uzbek_sentences_dataset")
label_list = ds["train"].features["label"].names   # 6 classes

MODEL = "bert-base-multilingual-cased"   # or "xlm-roberta-base"
tokenizer = AutoTokenizer.from_pretrained(MODEL)

def tokenize(batch):
    return tokenizer(batch["sentence"], truncation=True, padding="max_length", max_length=128)

tokenized = ds.map(tokenize, batched=True)

model = AutoModelForSequenceClassification.from_pretrained(
    MODEL,
    num_labels=len(label_list),
    id2label={i: l for i, l in enumerate(label_list)},
    label2id={l: i for i, l in enumerate(label_list)},
)

args = TrainingArguments(
    output_dir="./uzbek-classifier",
    num_train_epochs=5,
    per_device_train_batch_size=16,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
)

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized["train"],
    eval_dataset=tokenized["validation"],
)
trainer.train()

Limitations & known issues

• Size — 643 sentences is sufficient for experimentation but small for production-grade fine-tuning. Consider augmenting with mT5, back-translation, or additional scraping.
• Sources — both sources are news portals; the dataset does not cover conversational, literary, or scientific Uzbek.
• Script — Latin Uzbek only. Uzbek Cyrillic (still used in some contexts) and the Arabic script (used in historical texts) are not represented.
• Labeling noise — category labels are assigned heuristically (keyword matching + source feed prior). Some sentences may be mislabeled, particularly in the general class.
• Class imbalance — politics (21.6 %) is the largest class; general (12.3 %) is the smallest. Stratified splits preserve this ratio.

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Original articles are copyright of kun.uz and daryo.uz respectively. Sentences are reproduced here for non-commercial NLP research purposes under fair use.

Citation

If you use this dataset in your research, please cite it as:

@dataset{uzbek_sentences_2026,
  title     = {Uzbek News Sentences: A 6-class Topic Classification Dataset},
  year      = {2026},
  language  = {Uzbek (Latin script)},
  note      = {Scraped from kun.uz and daryo.uz}
}