Conflict-Aware Dual-Path Context Propagation for Multimodal Emotion Recognition in Conversation
Abstract
Multimodal Emotion Recognition in Conversation (MERC) aims to identify speakers' emotions by integrating conversational context with textual, acoustic, and visual cues. Existing methods remain limited in capturing emotion-relevant contextual dependencies and modeling cross-modal discrepancies during multimodal fusion. To address these challenges, we propose a Conflict-Aware Dual-Path Context Propagation framework (CDCP) for MERC. Specifically, CDCP first constructs dynamic semantic neighborhoods for each utterance via Adaptive Neighbor Selection, and then models contextual dependencies through speaker-aware Dual-Path Dynamic Multi-Hop Propagation to learn higher-order unimodal representations. To better handle cross-modal discrepancies, we further design a Conflict-Aware Expert Reconstruction module. It decomposes modality features into high- and low-frequency components in the frequency-domain, characterizes cross-modal consensus and modality-specific discrepancies through component interaction and fusion, captures potential cross-modal conflicts, and adaptively reconstructs emotion representations. Experiments on MELD and IEMOCAP demonstrate the effectiveness of CDCP, and further analyses validate the design of the proposed framework. Our code is available at https://github.com/MKMaS-GUET/CDCP.