BERT Roman Urdu Sentiment Analysis

Fine-tuned bert-base-multilingual-cased for 3-class sentiment (negative / neutral / positive) on Roman Urdu text.

Results (test set)

• Accuracy: 0.644
• Macro-F1: 0.632

Training

• Dataset: community-datasets/roman_urdu (20,229 sentences, 80/10/10 split)
• Epochs: 3, LR: 3e-5, batch size: 32, max length: 128

Limitations

Roman Urdu has no standard spelling, so spelling variants and slang can be misclassified. The model is biased towards the neutral class.