unicorn-s/nepal-border-sentiment-dataset

Nepal Border Sentiment Dataset

YouTube comments scraped from 17 Nepali news and commentary channels covering the 2026 Nepal–India border dispute, including the Prime Minister's parliamentary remarks. The dataset is labeled for 3-class sentiment (positive / neutral / negative).

Files

• —`nepal_border_comments.csv` — Raw scraped comments with channel and video URL
• —`nepal_comments_labelled.csv` — Translated and auto-labeled (silver standard) using cardiffnlp/twitter-xlm-roberta-base-sentiment
• —`comments_manual_review.csv` — 300 manually verified samples (gold standard, 100 per class)

Statistics

• —~2100 unique comments (after deduplication)
• —Languages: Nepali (Devanagari), Romanized Nepali, English, code-switched
• —300 manually verified gold-standard labels
• —~1800 silver-standard (auto-labeled) samples

How it was built

1. 1.Scraped via youtube-comment-downloader (no API key required)
2. 2.Translated to English using Google Translate via deep-translator
3. 3.Auto-labeled with multilingual XLM-RoBERTa sentiment model
4. 4.300 samples manually verified for the gold-standard split

Limitations

• —Single annotator for gold-standard labels (no inter-annotator agreement score)
• —Translation quality varies for Nepali slang and idioms
• —Class imbalance in auto-labels skewed toward negative sentiment
• —Sample focused on a specific political event — may not generalize broadly

Use & Citation

Published as a baseline for low-resource Nepali political sentiment NLP. Contributions welcome — see the GitHub repo for the full pipeline.