TajikNLPWorld/tajik-news-multilabel

Tajik News Multilabel Classification Dataset

Dataset Summary

This dataset contains 108,947 Tajik news articles annotated with multiple topic labels. Each document can belong to any subset of 14 predefined tags. The average number of labels per document is 5.27.

The labels were generated using a keyword‑based rule system that scans the article text for relevant terms. The dataset is intended for multilabel classification tasks, such as predicting all applicable topics for a given news article.

Uses

Direct Use

• Training multilabel text classification models (e.g., binary relevance, classifier chains, transformer‑based models).
• Evaluating tag prediction systems for Tajik news.
• Studying label co‑occurrence patterns.

Out-of-Scope Use

The dataset is not designed for single‑label classification; use the TajikNLPWorld/tajik-news-multiclass dataset for that purpose.

Dataset Structure

Data Fields

• content (string): Full text of the article (title + body, concatenated with \n\n)
• title (string): Article headline
• labels (list): List of tag names (e.g., ["Сиёсат", "Ҳукумат"])
• label_vector (list): Binary vector of length 14, one for each tag
• num_labels (int64): Number of tags assigned to this document
• original_category (string): The original normalized category (single) from the source
• 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

A single split (train) containing all 108,947 records. Users are encouraged to create their own train/validation/test splits.

Tags (14 labels)

Tag	Occurrences	% of documents
Ҳукумат	92,341	84.8%
Сиёсат	63,400	58.2%
Ҷаҳон	52,080	47.8%
Ҷамъият	50,069	46.0%
Осиёи Марказӣ	45,710	42.0%
Иқтисод	45,308	41.6%
Маориф	44,191	40.6%
Ҳуқуқ	42,101	38.6%
Дин	34,261	31.4%
Фарҳанг	31,893	29.3%
Ҳодиса	22,304	20.5%
Амният	19,598	18.0%
Варзиш	17,182	15.8%
Тандурустӣ	13,526	12.4%

Label Co‑occurrence (Top 10 Pairs)

Pair	Count	% of documents
Сиёсат + Ҳукумат	61,210	56.2%
Ҳукумат + Ҷаҳон	44,700	41.0%
Ҳукумат + Ҷамъият	44,522	40.9%
Иқтисод + Ҳукумат	41,057	37.7%
Осиёи Марказӣ + Ҳукумат	40,526	37.2%
Маориф + Ҳукумат	38,886	35.7%
Ҳукумат + Ҳуқуқ	38,878	35.7%
Сиёсат + Ҷаҳон	31,977	29.4%
Иқтисод + Сиёсат	31,011	28.5%
Сиёсат + Ҷамъият	30,513	28.0%

Dataset Creation

Curation Rationale

The dataset was created to enable multilabel classification in Tajik, a task that is more realistic for news articles because they often cover multiple topics simultaneously. Instead of relying solely on source categories (which are often single‑label), we derived multiple labels using a set of curated keywords for each tag.

Source Data

Data Collection and Processing

1. Collection: Articles were collected from eight Tajik news portals (see cluster dataset card for details) between 2015 and 2025.
2. Cleaning: HTML removed, whitespace normalized, deduplicated, filtered by length (50–10,000 characters).
3. Keyword‑based Labeling: For each of the 14 tags, a list of relevant Tajik keywords was manually compiled (e.g., for “Ҳукумат”: “ҳукумат”, “вазорат”, “президент”, etc.). If any keyword appeared in the article text (title + content), the tag was assigned. Tags were assigned independently; no manual validation was performed.
4. Final Format: Each article is stored with the original single category (from the source) and the computed multilabel vector.

Who are the source data producers?

The original content was produced by the news portals listed in the cluster dataset card.

Annotations

Annotations are rule‑based, not human‑annotated. This introduces noise: some tags may be assigned incorrectly if keywords appear in irrelevant contexts. However, the large scale and high label frequency make the dataset useful for training models that can learn from noisy labels.

Personal and Sensitive Information

No additional personal information was collected beyond what is already public in the articles.

Bias, Risks, and Limitations

• Label imbalance: Some tags (e.g., “Ҳукумат”) appear in 85% of documents, while others (e.g., “Тандурустӣ”) appear in only 12%. Models may predict frequent tags too readily.
• Keyword noise: The rule‑based approach may mislabel articles where keywords appear in non‑topical contexts (e.g., a sports article mentioning “government”).
• Correlation bias: The high co‑occurrence of certain tags (e.g., “Сиёсат” and “Ҳукумат”) may cause models to over‑predict them together.
• Source imbalance: The dataset is dominated by articles from asiaplus.tj (62%), which may affect generalization.

Recommendations

Users should:

• Use label‑weighted loss functions or oversampling to handle imbalance.
• Consider cleaning the labels by requiring multiple keyword occurrences or using more sophisticated rule‑based methods.
• Evaluate on a held‑out set and possibly collect a small human‑annotated test set for more accurate evaluation.

Citation

If you use this dataset, please cite:

@misc{arabov2025tajikmultilabel,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {Tajik News Multilabel Classification Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/TajikNLPWorld/tajik-news-multilabel}}
}

Dataset Card Authors

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

Dataset Card Contact

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