Conversational Emotion Recognition Using Self-Attention Mechanisms and Graph Neural Networks

Different from the emotion estimation in individual utterances, context-sensitive and speaker-sensitive dependences are vitally pivotal for conversational emotion analysis. In this paper, we propose a graph-based neural network to model these dependences. Specifically, our approach represents each utterance and each speaker as a node. To bridge the context-sensitive dependence, each utterance node has edges between immediate utterances from the same conversation. Meanwhile, the directed edges between each utterance node and its speaker node bridge the speaker-sensitive dependence. To verify the effectiveness of our strategy, we conduct experiments on the MELD dataset. Experimental results demonstrate that our method shows an absolute improvement of 1%~2% over state-of-the-art strategies.