Graph representation learning has emerged as a powerful paradigm for modeling complex non-Euclidean data by effectively capturing intricate relational patterns in graph-structured data, yet conventional heterogeneous graph neural networks commonly overlook the potential interactions between semantic and structural information. Recent advances in LLM-based graph representation learning offer new perspectives, where the fusion of large language models (LLMs) with graph neural networks (GNNs) simultaneously enhances both feature representation and structural modeling capabilities, significantly improving graph learning performance. We introduce a semantic passing mechanism that enables LLMs to replicate the GNN message-passing process in textual space through carefully designed prompts. We achieve LLM-guided neighbor aggregation through LLM-guided metapath selection and LLM-guided semantic passing processes. Our method effectively bridges structural and semantic information via language-based representations. The experimental results validate the efficacy of the proposed approach, offering a more generalized solution for heterogeneous graph learning tasks. The code is available at this URL.( \(^{1}\) https://github.com/zch65458525/LNAHGN