Skip to content

Inter-Dialog Contrastive Learning for Multimodal Emotion Recognition in Conversations

Oral presentation

Venue: IEEE ICASSP 2026
Author role: Second author, presenter
Year: 2026.05

Research approach

Selected neighboring representations across dialogues as positive pairs for audio-text contrastive alignment.

IDCL model architecture — IDCL model architecture
View full-size figure

Inter-dialog contrastive learning in IDCL — Inter-dialog contrastive learning in IDCL
View full-size figure

Results

With MELD pretraining, IEMOCAP weighted F1 is 85.8% for four classes and 66.6% for six classes.

All publications