Datasets:
license: mit
task_categories:
- text-classification
language:
- uk
- ru
tags:
- sentiment
- social
- networks
- Telegram
size_categories:
- 10K<n<100K
The corpus was constructed from multiple sources to ensure diversity and representation of real-world Ukrainian social discourse. We systematically scraped comments and posts from Ukrainian Telegram channels, collecting content dated between February 2022 and September 2024. The volume of the scraped documents amounts to 8,064 texts. Also, we integrated two publicly available datasets: TG samples from D. Baida (https://huggingface.co/datasets/dmytrobaida/autotrain-data-ukrainian-telegram-sentiment-analysis) with 3,000 samples and 1,000 Yakaboo book reviews (https://github.com/osyvokon/awesome-ukrainian-nlp). Furthermore, 1,000 product reviews from Hotline.ua were incorporated to diversify the content domains.
After deleting duplicates and boilerplate content, the final corpus included 12,224 texts covering various topics including politics, governmental services, entertainment, daily life, and consumer reviews. All texts in languages other than Ukrainian and Russian were filtered out due to their statistical insignificance.
Five annotators participated in the project, are native Ukrainian speakers with bilingual proficiency in Russian. The annotation guidelines instructed participants to classify texts according to four sentiment categories:
- Positive: posts containing expressions that reflect positive emotions (joy, support, admiration etc).
- Negative: texts containing expressions that reflect negative emotions (criticism, sarcasm, condemnation, aggression, doubt, fear etc).
- Neutral: documents where the author does not use either positive or negative expressions.
- Mixed: texts containing expressions from both positive and negative emotional spectra.