MoodShift — RoBERTa+ESA+TF-IDF+FL (Balanced 4500/class)
Novel contribution: LLM-based minority class augmentation via Groq (llama-3.3-70b) with self-consistency filtering using the trained model itself.
Results
| Model | Accuracy | Macro F1 |
|---|---|---|
| Original (imbalanced) | 0.9250 | 0.8849 |
| LLM-Augmented (4500/class) | 0.9285 | 0.8893 |
Key Improvements
- Love F1: 0.8344 → 0.8613 (+0.0269)
- Surprise F1: 0.7521 → 0.7576 (+0.0054)
ICCA 2026 HCI Research — MoodShift Adaptive Chatbot
- Downloads last month
- 72