Effectively fusing heterogeneous semantic information while ensuring training stability remains a critical challenge in graph learning. This paper proposes HGNN-SF, a robust architecture designed for link prediction in dense and complex heterogeneous graphs. The core of the proposed model is a novel Semantic Fusion Layer (SF-Layer) that integrates a two-level attention mechanism. First, Intra-Type Spectral Attention incorporates spectral normalization into relation-specific attention modules to stabilize message propagation and prevent gradient explosion in high-degree, dense graph structures. Second, a Hierarchical Semantic Fusion mechanism dynamically assigns importance weights to aggregated messages from distinct relation types, enabling effective semantic integration without relying on predefined meta-paths. To regulate information flow and mitigate over-smoothing in deep heterogeneous message passing, node representations are updated using a Gated Recurrent Unit (GRU). An ablation study demonstrates that this gated update significantly outperforms conventional residual connections in preserving feature discriminability. Additionally, Hard Negative Sampling is employed during training to enhance model robustness against structurally challenging negative instances. Extensive experiments conducted on four large-scale, real-world datasets (DBLP, ACM, Amazon, and LiveJournal) demonstrate that HGNN-SF consistently outperforms strong graph neural network (GNN) and heterogeneous GNN (HGNN) baselines in link prediction tasks. Notably, on the highly challenging LiveJournal dataset, HGNN-SF achieves a Test AUC of 0.9044, confirming the efficacy of combining spectral attention and gated semantic fusion for stable, accurate representation learning in heterogeneous networks.