Cross-Domain Recommendation can significantly mitigate the challenges posed by sparse data for recommendation systems. Relevant studies indicate domain-specific preferences negatively impact the recommendation performance of target domains based on domain-shared information. Recent research considers domain-invariant and domain-specific features. Nevertheless, these intricately entangled features are hardly discerned for differentiation and the semantic diversity in heterogeneous relationships tends to be understated. In light of this, a novel model entitled Disentangled Representations for Cross-Domain Recommendation via Heterogeneous Graph Contrastive Learning (DHCL) is proposed. We derive domain-invariant and domain-specific representations, capturing both commonalities and unique features across diverse domains. Heterogeneous graph and meta-path are used to assist in enhancing the amount of information. We formulate dual contrastive learning tasks to further obtain optimal disentangled representations. Comprehensive experiments on three pairs of authentic review datasets highlight the superiority of DHCL over SOTA recommendation methods.