Datasets:

uznlp-uz
/

uz_proverb

language:
  - uz
license: cc-by-4.0
task_categories:
  - text-classification
  - text-generation
  - token-classification
pretty_name: Uzbek Proverbs Dataset (8.5K)
size_categories:
  - 1K<n<10K

Uzbek Proverbs Dataset (UZBEKPROVERBS-8.5K)

📌 Description

Proverbs are culturally salient figurative expressions, yet Uzbek still lacks standardized NLP resources for their automatic identification. This dataset introduces UzbekProverbs-8.5K, a curated, machine-readable proverb resource for Uzbek, derived from a source inventory of 8,514 entries.

The resource contains:

8,477 unique proverb strings
2,998 glossed entries
70 normalized thematic groups

A structured curation pipeline was applied, including:

label normalization
duplicate resolution
canonical identifier assignment
metadata standardization

These steps address inconsistencies typical in raw proverb collections, such as orthographic variation, uneven annotation coverage, and polythematic duplication.

🧠 Benchmark Definition

This dataset is released as a benchmark for automatic proverb identification in Uzbek corpora, supporting reproducible research in low-resource figurative language processing.

Supported Tasks

1. Proverb Detection (Classification)

Input: sentence
Output: binary label (contains proverb / does not contain proverb)

2. Proverb Localization (Span Detection)

Input: sentence
Output: span of proverb text within the sentence

3. Canonical Proverb Linking

Input: proverb variant (possibly noisy or inflected form)
Output: canonical proverb entry

📊 Dataset Structure

The dataset is stored in TSV format (tab-separated) with the following columns:

Column name	Description
`id`	Unique identifier
`Maqollar`	Uzbek proverb text
`Maqollarning guruhlari`	Semantic category
`Maqollarning ma'nolari`	Explanation / meaning

🧾 Example

{
  "id": 1,
  "Maqollar": "Ha’ga 'Hu' kelar.",
  "Maqollarning guruhlari": "Yaxshilik va yomonlik haqida maqollar",
  "Maqollarning ma'nolari": "Yaxshilik yoki yaxshi muomala odatda javobsiz qolmaydi, unga munosib javob qaytadi."
}

🏷 Label Schema

The dataset includes normalized semantic categories such as:

Yaxshilik va yomonlik
Donolik va nodonlik
Mehnatsevarlik va dangasalik
Do‘stlik va dushmanlik
Sabrlilik va sabrsizlik
Adolat va insofsizlik
Yaxshi so‘z va yomon so‘z

📏 Evaluation Protocol

The benchmark supports multiple evaluation settings:

Classification (Detection)

Accuracy
Precision / Recall / F1-score

Span Detection (Localization)

Exact Match (EM)
Token-level F1

Canonical Linking

Top-1 accuracy
Retrieval-based metrics (MRR, Recall@k)

📊 Data Splits

Note: predefined splits may be added in future versions.

Recommended split strategy:

Train: 80%
Validation: 10%
Test: 10%

🎯 Use Cases

This dataset can be used for:

🏷 Proverb classification and detection
🧠 Semantic similarity and retrieval
🤖 Fine-tuning Uzbek language models (BERT, LLaMA, Qwen, etc.)
🔍 Information retrieval systems
📚 Educational and linguistic research tools

⚙️ Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("ruhilloalaev/proverb_Uzbek")
print(dataset["train"][0])

⚠️ Notes

Data is in Uzbek (Latin script)
File uses tab (\t) as delimiter
Proverbs may exhibit:
- morphological variation
- orthographic inconsistency
- figurative ambiguity

📜 License

This dataset is released under the CC-BY-4.0 License.