Context-Aware Dynamic Graph Learning for Multimodal Emotion Recognition with Missing Modalities

The missing modality problem in multimodal emotion recognition in conversations (MERC) remains a key challenge in real-world applications, as existing methods often fail to preserve semantic coherence or achieve robust cross-modal reconstruction. To address this issue, we propose a unified framework that combines adaptive dialogue graphs, Large Language Models (LLMs)-guided emotional context extraction, and cross-modal semantic alignment. Our approach leverages dynamic graph ordinary differential equations to model temporal and speaker-specific dynamics, while LLMs refine textual representations by extracting emotion-relevant keywords conditioned on conversational context. An additional alignment loss enforces semantic consistency between reconstructed and observed modalities. Experiments on IEMOCAP and MELD demonstrate that our method achieves state-of-the-art performance under both complete- and missing-modality conditions, outperforming prior approaches and showing strong robustness to incomplete inputs. Code is available at https://github.com/premiree/CDAGL.git.