Mutual Attention Network for Multi-label Emotion Recognition with Graph-Structured Label Representations

Multi-label emotion classification (MLEC) tasks have gained significant attention due to their closer alignment with real-life emotional expressions. However, previous studies mostly focused on exploring semantic information in texts, while often neglecting the emotional information encoded in the labels. To address this, an advanced framework called the Text-Label Mutual Attention Network(TLMAN) model is proposed to thoroughly investigate the complex interactions between text and labels. The model effectively combines a text representation learning module, a label representation learning module and a mutual attention module to delve into the intricate relationships between labels and textual content, focusing on a holistic understanding of emotional information integration. To enhance MLEC methods, this study integrates Graph Convolutional Networks (GCN) to reveal label dependencies. The TLMAN model achieved AP scores of 76.93% and 77.38% on the Ren-CECps and NLPCC2018 datasets, significantly improving the performance of MLEC tasks.