Triple disentangled representation learning for multimodal affective analysis.