unicorn-s/nepal-border-sentiment-xlmr

Nepal Border Sentiment Classifier

XLM-RoBERTa-base fine-tuned for 3-class sentiment classification (positive/neutral/negative) on Nepali and English YouTube comments discussing the Nepal–India border dispute.

Performance

• Accuracy: ~50% on gold-standard evaluation set (manually verified)
• Random baseline: 33% (3-class)
• F1 (weighted): ~0.50

Dataset

Trained on the Nepal Border Sentiment Dataset — ~2100 YouTube comments scraped from 17 Nepali news and commentary channels.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained("unicorn-s/nepal-border-sentiment-xlmr")
tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")

text = "Your Nepali or English comment here"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
labels = ['negative', 'neutral', 'positive']
print({labels[i]: float(probs[0][i]) for i in range(3)})

Limitations

• Small training set (~1800 silver-labeled samples)
• Single annotator for gold-standard evaluation
• Translation quality varies for Nepali slang and code-switched text
• Class imbalance in training data skewed toward negative sentiment

Citation / Contribution

This is published as a baseline for low-resource Nepali political sentiment analysis. Contributions of additional labeled data and improved baselines are welcome — see the GitHub repo.