TajikNLPWorld/tajik-news-multiclass

Tajik News Multiclass Classification Dataset

Dataset Summary

This dataset contains 97,531 Tajik news articles labeled with one of 12 topic categories. It is designed for multiclass text classification tasks. The categories are well-balanced and cover major news topics such as Ҷамъият (Society), Иқтисод (Economy), Сиёсат (Politics), Ҳодиса (Incidents), and others.

The data have been cleaned, deduplicated, and filtered to ensure consistent quality. Articles with empty or uninformative categories were excluded.

Uses

Direct Use

The dataset can be used for:

• Supervised multiclass classification of Tajik news texts
• Benchmarking Tajik language models
• Training classifiers for news topic detection
• Educational purposes in NLP for low-resource languages

Out-of-Scope Use

The dataset is not intended for clustering or unsupervised learning (a separate dataset is available for that: TajikNLPWorld/tajik-news-cluster). It also does not contain multiple labels per document (see the multilabel dataset for that).

Dataset Structure

Data Fields

• content (string): Full text of the article (title + body, concatenated with \n\n)
• title (string): Article headline
• category (string): One of the 12 topic categories
• content_length (int64): Length of content in characters
• resource (string): URL of the article (if available)
• date (string): Publication date (if available; otherwise empty)

Data Splits

The dataset contains a single split (train) with all 97,531 records. For evaluation, users should create their own train/validation/test splits, e.g., 80/10/10.

Categories (12 classes)

Category	Count	Percentage
Ҷамъият	25,448	26.1%
Иқтисод	13,833	14.2%
Сиёсат	12,467	12.8%
Ҳодиса	9,597	9.8%
Амният	8,599	8.8%
Хабарҳо	8,406	8.6%
Фарҳанг	5,424	5.6%
Варзиш	4,240	4.3%
Ҳуқуқ	2,755	2.8%
Ҳукумат	2,653	2.7%
Ҷаҳон	2,489	2.6%
Маориф	1,620	1.7%

Dataset Creation

Curation Rationale

The dataset was created to provide a standardized benchmark for multiclass classification in Tajik, a low-resource language. It aggregates articles from multiple sources and applies a consistent category normalization to reduce noise.

Source Data

Data Collection and Processing

1. Collection: Articles were collected from eight Tajik news portals (see the cluster dataset card for full list) between 2015 and 2025.
2. Cleaning: Removed HTML, extra whitespace, deduplicated, filtered by length (50–10,000 characters).
3. Category Normalization:
   • Original categories (e.g., tajikistan/politics, Хабарҳо ва рӯйдодҳо, Iqtisod) were mapped to a unified set of 12 classes using a rule‑based dictionary and keyword matching.
   • Articles with empty or garbage categories were discarded.
4. Final Format: Each article is stored as a JSON object with the fields above.

Who are the source data producers?

The original content was produced by the respective news portals listed in the cluster dataset card. The dataset is based on publicly available news articles.

Annotations

Categories were derived from the original category field provided by the sources, supplemented with keyword‑based rules. No manual annotation was performed, but the normalization rules were carefully designed to map similar concepts to the same class.

Personal and Sensitive Information

The dataset contains news articles that may mention public figures, but no attempt has been made to remove personal information. All data are already publicly available.

Bias, Risks, and Limitations

• Class imbalance: The smallest class (Маориф) has only 1.6% of the data. Models may underperform on minority classes unless weighting or oversampling is used.
• Source imbalance: The dataset is dominated by asiaplus.tj (62% of records), which may affect generalization to other styles.
• Date coverage: Not all articles have a reliable date field.
• Language: Tajik (Cyrillic) only.

Recommendations

Users should:

• Use stratified splits to preserve class distribution.
• Consider class weights or resampling when training.
• Evaluate on a held‑out set from multiple sources.

Citation

If you use this dataset, please cite:

@misc{arabov2025tajikmulticlass,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {Tajik News Multiclass Classification Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/TajikNLPWorld/tajik-news-multiclass}}
}

Dataset Card Authors

• Arabov Mullosharaf Kurbonovich
  Email: cool.araby@gmail.com

Dataset Card Contact

For questions, please contact cool.araby@gmail.com.