Multimodal Data Fusion for E-commerce Dialogue Intent Recognition

With the rapid development of e-commerce platforms, multimodal data fusion (text, images, etc.) for dialogue intent recognition has emerged as a research hotspot. To address the interference from redundant visual features, this paper proposes VD-BERT (Region-Aware Visual Dynamic BERT), an enhanced VisualBert-based model for e-commerce intent recognition. Before cross-modal fusion, visual features undergo preprocessing through RA-ViT (Region-Aware Visual Transformer), whose output serves as visual input parameters for multimodal fusion. Furthermore, we introduce DynMM (Dynamic Multimodal Mixer) to optimize feature fusion and reasoning efficiency. Experimental results demonstrate that our VD-BERT model achieves superior accuracy over existing multimodal approaches (VisualBert, ViLT, CLIP, BLIP) on the MMDIS-2025 dataset.