Multimodal Emotion Recognition Based on Global Information Fusion in Conversations
Oral presentation
Research approach
Used a Multimodal Transformer to fuse long-range context across speech, video and text. An RGCN models within-speaker, between-speaker and cross-modal relationships for utterance-level emotion classification.
Results
Six-class IEMOCAP evaluation: accuracy 68.21% and weighted F1 68.17%, compared with weighted F1 of 67.60% for COGMEN.
All publications