BiM-MFGCN: a bidirectional Mamba and multi-frequency graph convolutional network for multimodal emotion recognition in conversation

Abstract

Emotion recognition in conversation (ERC) automatically identifies the emotional states conveyed by individual utterances in multi-turn dialogues and is valuable for intelligent customer service, mental-health monitoring, and human-computer interaction. Existing multimodal ERC methods still face challenges in capturing bidirectional contextual dependencies, preserving local emotional variations, and integrating complementary cross-modal cues. To address these issues, we propose BiM-MFGCN, a bidirectional Mamba and multi-frequency graph convolutional network for multimodal ERC. The model contains a bidirectional Mamba temporal modeling module (BAM) and a multi-frequency graph attention module (FGA). BAM employs forward and backward Mamba2 paths to encode preceding and subsequent dialogue context. FGA constructs an adaptive composite utterance graph, separates low-, mid-, and high-frequency graph-spectral components, and performs cross-modal interaction on frequency-enhanced representations. Experiments on MELD and IEMOCAP show that BiM-MFGCN achieves 66.66% accuracy and 67.82% weighted F1 on MELD, and 70.11% accuracy and 70.12% weighted F1 on IEMOCAP, demonstrating its effectiveness for multimodal ERC.