RoBERTa Multi-Label Emotion Classifier
Fine-tuned roberta-base on SemEval-2018 Task 1, Subtask E-c for 11-label emotion classification.
Labels
anger 路 anticipation 路 disgust 路 fear 路 joy 路 love 路 optimism 路 pessimism 路 sadness 路 surprise 路 trust
Performance (Test Set)
| Metric | Score |
|---|---|
| Macro-F1 | 58.4% |
| Micro-F1 | 68.9% |
| Jaccard | 54.7% |
Training
- Architecture: RoBERTa-base + linear classification head
- Loss: BCEWithLogitsLoss (class-weighted)
- Optimizer: AdamW (lr=2e-5, wd=0.01)
- Epochs: 5 with linear warmup