README.md

metadata

language:
  - en
license: cc-by-4.0
task_categories:
  - text-classification
task_ids:
  - topic-classification
  - sentiment-classification
tags:
  - linguistics
  - ukraine
  - toponyms
  - language-policy
  - kyivnotkiev
size_categories:
  - 10K<n<100K

KyivNotKiev Computational Linguistics Corpus

A balanced, labeled corpus of texts containing Ukrainian and Russian toponym variants (e.g., "Kyiv" vs "Kiev"), annotated with context categories and sentiment.

Dataset Description

• Curated by: Ivan Dobrovolskyi
• Language: Primarily English
• License: CC-BY 4.0
• Paper: #KyivNotKiev: A Large-Scale Computational Study of Ukrainian Toponym Adoption (forthcoming)
• Website: https://kyivnotkiev.org

Dataset Summary

29,938 texts across 55 Ukrainian-Russian toponym pairs from 4 sources (Reddit, YouTube, GDELT news articles). Each text is labeled with:

• Context category: politics, war_conflict, sports, culture_arts, food_cuisine, travel_tourism, academic_science, history, business_economy, general_news
• Sentiment: positive, neutral, negative
• Variant: which toponym form (russian/ukrainian) appears in the text

Intended Uses

• Studying language policy adoption in media and social platforms
• Training toponym context classifiers
• Analyzing sentiment differences between spelling variants
• Cross-source and temporal analysis of naming conventions

Dataset Structure

Data Fields

• pair_id: Integer ID of the toponym pair
• text: The full text content
• variant: "russian" or "ukrainian" — which spelling form appears
• source: Data source (reddit, youtube, gdelt)
• year: Publication year
• context_label: Annotated context category
• context_confidence: Annotation confidence (0-1)
• sentiment_label: Sentiment annotation
• sentiment_score: Sentiment score (-1 to 1)
• word_count: Number of words in text
• matched_term: The specific toponym form found in text

Splits

Split	Count
train	23,950
validation	2,993
test	2,993

Balance Report

See balance_report.json for detailed per-pair, per-source, per-variant distributions and documented shortfalls.

Collection Methodology

1. Reddit: Titles and bodies from Arctic Shift API + Reddit search (2010-2026)
2. YouTube: Video titles and descriptions via yt-dlp (2010-2026)
3. GDELT: News article bodies fetched from URLs using trafilatura (2010-2026)
4. Balancing: Stratified sampling by pair × source × variant × year stratum
5. Annotation: Llama 3.1 70B-Instruct with human validation on 200 random samples
6. Fetch transparency: All GDELT URL fetch attempts logged in fetch_log.parquet

Citation

@article{dobrovolskyi2026kyivnotkiev,
  title={#KyivNotKiev: A Large-Scale Computational Study of Ukrainian Toponym Adoption},
  author={Dobrovolskyi, Ivan},
  year={2026}
}