Visually Rich Document (VRD) presents a significant challenge for information extraction in the field of document intelligence due to their complex and diverse layouts and the inherent difficulty of multi-modal feature fusion. Existing methods can be broadly categorized into two groups. First, conventional large-scale pre-trained models, while possessing strong feature representation capabilities, have limited capacity for spatial structure modeling, making it difficult to handle complex layouts and irregular document structures effectively. Second, graph-based methods can explicitly model relationships between regions but often rely on manual graph construction and hand-crafted feature design, which results in poor generalization and heightened sensitivity to structural variations and noise. To address these limitations, this paper proposes an innovative multi-modal feature fusion network, TriBiaNet, which achieves efficient integration of visual and textual information through baffine structural relationship modeling. Specifically, we propose a Dynamic Condition Aware Enhancer (DCAE) that adaptively produces cross-modal weight assignments, enhancing the flexibility of feature fusion. Additionally, we propose a Bidirectional Cross-modal Attention Module (BCAM) to enable efficient and compact multi-modal feature interaction, ensuring robust performance even under limited computational resources. Finally, we propose a Collaborative Prediction Module (CPM) that combines the hierarchical attention convolution mechanism with dual-affine relationship modeling, significantly strengthening entity recognition and relation reasoning to improve information extraction effectiveness. Experimental results demonstrate that TriBiaNet achieves superior performance compared to state-of-the-art methods.