The paper introduces a novel framework called GTC (GNN-Transformer Co-contrastive Learning) for self-supervised heterogeneous graph representation. GTC combines the strengths of Graph Neural Networks (GNNs) and Transformers to address the over-smoothing issue in GNNs and to capture multi-hop neighbor information effectively. GNNs are known for their ability to aggregate local information, while Transformers excel at modeling global information and multi-hop interactions. GTC leverages both GNN and Transformer branches to encode node information from different views and establishes contrastive learning tasks based on the encoded cross-view information. The Transformer branch includes Metapath-aware Hop2Token and CG-HetPhormer, which can cooperate with GNN to attentively encode neighborhood information from different levels. The proposed method is evaluated on real heterogeneous graph datasets, showing superior performance compared to state-of-the-art methods. The experiments demonstrate that GTC can maintain high performance and stability even when the model is deeper, proving its effectiveness in capturing multi-hop neighbor information without over-smoothing.The paper introduces a novel framework called GTC (GNN-Transformer Co-contrastive Learning) for self-supervised heterogeneous graph representation. GTC combines the strengths of Graph Neural Networks (GNNs) and Transformers to address the over-smoothing issue in GNNs and to capture multi-hop neighbor information effectively. GNNs are known for their ability to aggregate local information, while Transformers excel at modeling global information and multi-hop interactions. GTC leverages both GNN and Transformer branches to encode node information from different views and establishes contrastive learning tasks based on the encoded cross-view information. The Transformer branch includes Metapath-aware Hop2Token and CG-HetPhormer, which can cooperate with GNN to attentively encode neighborhood information from different levels. The proposed method is evaluated on real heterogeneous graph datasets, showing superior performance compared to state-of-the-art methods. The experiments demonstrate that GTC can maintain high performance and stability even when the model is deeper, proving its effectiveness in capturing multi-hop neighbor information without over-smoothing.