Heterogeneous Information Networks (HINs) present complex data relationships due to their diverse node and edge types, posing challenges for effective representation learning. While Graph Transformers (GTs) have shown promising results, their performance is heavily dependent on large amounts of labeled data, which is often expensive and difficult to obtain. To mitigate this issue, we introduce Contrastive Learning for Heterogeneous Graph Transformers (CL-HGT), a self-supervised learning framework designed to enhance the learning of structural and heterogeneous information in label-scarce HINs. CL-HGT integrates a dual-view contrastive learning objective that combines global relational and local structural encodings, enabling Graph Transformers to better capture both long-range dependencies and fine-grained structural information. Additionally, we propose Heterogeneous Segment Mixup (HS-Mixup), a novel data augmentation technique tailored to tokenized graph inputs, which balances semantic consistency and diversity in augmented data. Experimental results demonstrate that CL-HGT outperforms baseline models across multiple HIN benchmarks, improving generalization and robustness in label-scarce scenarios.