Sinhala Emotion Recognition

This project addresses the lack of emotional NLP tools for Sinhala by introducing a manually annotated dataset of 10,000+ social media comments[cite: 54, 257]. I engineered a hybrid SinBERT-CNN model to extract contextual embeddings alongside local syntactic n-grams. The research involved a rigorous ablation study which revealed that while hybrid models excel in English, the agglutinative nature of Sinhala and sub-word tokenization fragmentation can lead to 'destructive interference'. The final system provides a benchmark Macro F1-score of ~0.71, establishing a baseline for future multi-class emotion detection in low-resource South Asian languages.