Self-Attentive Feature-Level Fusion for Multimodal Emotion Detection

Multimodal emotion recognition is the task of detecting emotions present in user-generated multimedia content. Such resources contain complementary information in multiple modalities. A stiff challenge often faced is the complexity associated with feature-level fusion of these heterogeneous modes. In this paper, we propose a new feature-level fusion method based on self-attention mechanism. We also compare it with traditional fusion methods such as concatenation, outer-product, etc. Analyzed using textual and speech (audio) modalities, our results suggest that the proposed fusion method outperforms others in the context of utterance-level emotion recognition in videos.