Graph Attention Network With Collaborative Learning for Conversational Emotion Recognition

Emotion recognition in conversations (ERC), a typical cyber-physical-social system (CPSS) application, is becoming increasingly popular in smart homes. Nowadays, many studies have developed algorithms for this purpose. However, they face two main challenges, i.e., not considering relevance in graph feature fusion and under utilizing relationships between dialogues. To address this, we propose a framework using graph attention network with collaborative learning for ERC. It is based on multiview attention and multiscale supervision. Specifically, we transform conversation data into a graph structure. Then, this graph matrix goes through a graph attention branch to integrate global and local information. The output is processed by the multiscale attention branch to extract time-related features, which are then fed into a fully connected layer to determine the speaker’s emotions. Notably, we also design a multiscale supervision strategy to enhance learning effectiveness. Our framework is trained and tested on four typical datasets, i.e., IEMOCAP, MELD, DailyDialog, and EmoryNLP. Experimental results well determined that it is effective and has advantages over peer state-of- the-art methods.