metadata
language: en
tags:
- text-classification
- emotion-detection
- mental-health
- hci
- distilbert
- cnn
datasets:
- go_emotions
metrics:
- f1
- accuracy
MoodShift — Hybrid DistilBERT + CNN Emotion Classifier
Novel hybrid architecture for HCI research (ICCA 2026).
Architecture
- DistilBERT → [CLS] global context (768-dim)
- CNN (kernels 2,3,4) → local n-gram emotion patterns (384-dim)
- Fusion → 1152-dim → Linear(512) → Linear(7)
Labels
positive | sadness | anger | anxiety | confusion | curiosity | neutral
Best Val F1: 0.9141