<scp>CSG</scp> ‐Net: A Causal‐Aware Structural Graph Network for Multimodal Music Emotion Recognition

ABSTRACT Multimodal Music Emotion Recognition (MMER) aims to bridge the ‘affective semantic gap’ between machine perception and human emotion by fusing information from audio and lyrics. However, state‐of‐the‐art models encounter a performance plateau, a problem we trace back to three persistent, structural challenges: (1) the neglect of music's intrinsic narrative structure, where models predominantly treat music as linear feature sequences; (2) a biased dependency on certain modalities, especially the ‘Lyrical Dominance Bias’, where models over‐rely on explicit lyrical cues, particularly in cases of emotional conflict; and (3) data modality incompatibility, where existing frameworks struggle to process both lyrical songs and instrumental pieces uniformly. To systematically address these issues, we introduce the Causal‐aware Structural Graph Network (CSG‐Net). CSG‐Net contributes to the field by uniquely integrating structural and causal priors into a unified MMER framework. First, we propose an adaptive structural topology module to parameterise the hierarchical logic of emotional progression and repetition, thereby moving beyond the limitations of linear modelling. Second, we pioneer the use of causal inference in this domain, designing a training and inference mechanism based on counterfactual intervention. This allows the model to disentangle and suppress the direct, often biasing, influence of lyrics, forcing it to learn the more robust causal relationship between audio and text. Finally, through an efficient random lyric masking strategy, CSG‐Net achieves seamless compatibility with different musical forms within a single model. Extensive experiments demonstrate that CSG‐Net significantly outperforms 14 highly competitive baselines, establishing a new state‐of‐the‐art specifically on the MER‐500 and MuSe‐CaR benchmarks under standard evaluation protocols, achieving an absolute accuracy improvement of +2.7% on MER‐500 and +2.2% on MuSe‐CaR ( p &lt; 0.005). Notably, the model limits performance degradation to only 12% during zero‐shot transfer to purely instrumental data (compared to &gt; 22% in baselines), providing consistent quantitative evidence of its superior cross‐modal generalisation and robustness.