UAReviews / README.md
woters's picture
Create README.md
161c772 verified
metadata
language:
  - uk
license: cc-by-4.0
tags:
  - sentiment-analysis
  - emotion-detection
  - intent-classification
  - ukrainian
  - benchmark
task_categories:
  - text-classification
pretty_name: UAReviews

UAReviews: Ukrainian Emotion and Intent Benchmark (v1.0)

UAReviews is a curated benchmark of 11 580 Ukrainian user reviews and feedback comments labeled for both emotion and intent category.
It is designed for evaluating and fine-tuning sentiment, emotion, and intent-understanding models for the Ukrainian language.


Highlights

  • 7-class emotion and 5-class intent category annotation schema
  • Public sector reviews were kindly provided by the Ministry of Digital Transformation of Ukraine
  • A little re-balanced with a small, re-labeled subset of COSMUS to improve underrepresented categories

Data Composition

Source Description
Ministry of Digital Transformation dataset Original user reviews and service feedback from the Diia ecosystem and related municipal channels
COSMUS subset Selected and re-labeled samples used to balance low-frequency categories (Question / Request for Help, Suggestion / Idea, Neutral Comment)

All COSMUS records were re-labeled with the unified UAReviews schema (and marked with COSMUS source).
Each record retains its origin under the source field.


Dataset Structure

Field Description
id Unique identifier
content Original Ukrainian text
rating 1–5 star rating, if available
source "original" or "cosmus"
final_emotion One of: Happiness, Sadness, Anger, Fear, Disgust, Surprise, Neutral
final_category One of: Gratitude / Positive Feedback, Complaint / Dissatisfaction, Question / Request for Help, Suggestion / Idea, Neutral Comment
split "train", "test", "challenge"
length Character count of content

Statistics

Metric Value
Samples 11 580
Average text length 148 characters

Category distribution

Category Count Share
Gratitude / Positive Feedback 7 440 64 %
Complaint / Dissatisfaction 2 730 24 %
Question / Request for Help 615 5 %
Neutral Comment 418 4 %
Suggestion / Idea 377 3 %

Emotion distribution

Emotion Count Share
Happiness 7 557 (65%)
Anger 2 264 (20%)
Neutral 1 117 (10%)
Sadness 424 (4%)
Disgust 106 (0.9%)
Surprise 57 (0.5%)
Fear 55 (0.5%)

License

CC BY 4.0 — free to use, modify, and redistribute with attribution.
Portions derived from COSMUS are released under the same license.


Acknowledgments

Developed by KSE NLP Lab at the Kyiv School of Economics,
in collaboration with the Ministry of Digital Transformation of Ukraine.

We thank Y. Shynkarov et al. for the open COSMUS dataset that supported category balancing.