Improved Multi-hop Reasoning Through Sampling and Aggregating

Multi-hop reasoning over is helpful in knowledge graphs to unearth intricate associations and implicit information among entities through multi-step reasoning jumps. However, existing approaches still face the challenges of noise and sparsity. This is due to the fact that this issue it is difficult to identify head and tail entities along long and complex paths. To address this issue, we propose a novel multi-hop reasoning model based on Dual Sampling strategies and aggregation of Multi-Relational Types of entities and relationships, named DSMRT in short. In particular, the dual sampling strategy to address noise and sparsity of training data entities, including the forward traversal and the backward verification. Afterward, we distinguish the semantic types of different entities and relationships by constructing their type perception representations. Extensive experiments demonstrate that the proposed DSMRT model can adeptly oversee the sampling process, ensuring both balance and representativeness of the data. Additionally, it successfully mitigates challenges like noise and information gaps through the judicious application of type information.