In this paper, we address the practical yet challenging task of emotion recognition (ERC) in conversation, and propose a novel multimodal approach, the short-range graph neural network (LSDGNN). Based on the directed acyclic graph (DAG), we construct a long-range graph neural network and a short-range graph neural network to obtain multimodal features of distant and adjacent utterances, respectively. We use a differential regularizer to make long-range and short-range features as distinct as possible in expressions while allowing mutual influence between the two modules, and integrate bilinear modules to facilitate feature interaction. We also propose an improved curriculum learning (ICL) to address the data imbalance problem. We design a “weighted emotion change” index that emphasizes changes in similar emotions by calculating the similarity between different emotions, and develop a difficulty measure to enable a learning process that learns easy samples first. Experimental results on the IEMOCAP and MELD datasets show that the proposed model outperforms existing benchmarks.
