Inferring Speaking Styles from Multi-modal Conversational Context by Multi-scale Relational Graph Convolutional Networks.