Heterogeneous information networks effectively model complex real-world systems with diverse entity types and relationships. However, many practical scenarios face challenges where type information becomes unavailable or unreliable, resulting in Latent Heterogeneous Graphs (LHGs). This absence of explicit type information renders traditional heterogeneous graph methods inapplicable. To address this challenge, we introduce Path-Aware Graph Neural Network (PA-GNN), a novel architecture designed specifically for link prediction in LHGs. PA-GNN leverages innovative techniques including attention-based path representations, context-aware neighbor modulation, and multi-level attention aggregation to effectively capture latent semantics in heterogeneous structures without relying on explicit type information. Through extensive experiments on three benchmark datasets—FB15k-237, WN18RR, and DBLP—with their type information deliberately hidden, we demonstrate that PA-GNN consistently outperforms state-of-the-art methods. Our comprehensive ablation studies further verify the effectiveness of each proposed component and illuminate how PA-GNN successfully captures latent semantics in heterogeneous structures. The superior performance of our model provides valuable insights for representation learning in scenarios where type information is unavailable or incomplete.