<p>Relation prediction is a critical approach to addressing the incompleteness of knowledge graphs. Research has shown that learning long-term semantic dependency information from multi-hop neighbourhoods is essential for fully capturing the contextual semantics of knowledge graphs. Based on the complex graph structure of knowledge graphs, multi-hop neighbourhood structures include relational paths between entities and the multi-level local neighbourhoods of given entities. However, existing studies often focus on single-dimensional learning, making it prone to overlook specific patterns when predicting missing relations. Prior research relies on static path aggregation or sequence encoding methods, particularly in learning the multi-hop structures of relational paths, neglecting the dynamic interactions between relational paths.In this paper, we propose a model named <i>R-APSC</i>, which aims to learn contextual representations of knowledge graphs from two types of multi-hop topological structures: (i) by constructing relationship path subgraphs to characterize sets of relation paths between entities and proposing an attention-enhanced recursive message-passing scheme to capture the graph structure within relation paths and their dynamic interactions; and (ii) by integrating the local multi-hop neighbourhood information of entities to capture their intrinsic properties and deep contextual semantics. Experimental results demonstrate that the <i>R-APSC</i> model significantly outperforms state-of-the-art relation prediction models on multiple benchmark datasets, validating the importance of understanding relation semantic propagation and dependencies from a global perspective.</p>