<p>Heterogeneous Graphs (HGs) naturally capture complex interactions across various domains. However, representation learning on real-world HGs is often hindered by missing node attributes. The widespread missing node attributes significantly compromise the performance of standard feature-dependent Graph Neural Networks (GNNs). This challenge is compounded by network heterogeneity: while relying on local topology to complete attributes is effective in relational domains where neighbor attributes directly mirror the target node’s semantics, it becomes unreliable when neighbor types significantly differ from the target node in semantics. To address this, we propose HINAC, an end-to-end Graph Transformer framework augmented by Large Language Models (LLMs) that jointly tackles semantic recovery and heterogeneous structure learning. HINAC introduces an LLM-Augmented Attribute Completion (LAAC) module. Rather than relying entirely on the graph, LAAC extracts missing semantics from open-world textual knowledge and explicitly aligns them with local topologies via a meta-path-free, heterogeneous attention mechanism, providing a reliable semantic foundation. Taking these enriched semantics as input, a Message-Aware Context Encoder (MACE) and a Heterogeneity-Based Distribution Encoder (HADE) capture local structural dependencies and provide a probabilistic view of heterogeneous neighborhoods, respectively. Finally, a node-wise Transformer leverages these encoded representations to explicitly capture long-range interactions. Extensive experiments on multiple benchmark datasets demonstrate that HINAC accurately completes missing attributes and establishes superior performance on downstream tasks including node classification and clustering.</p>