A Novel Concept-Driven Negative Sampling Mechanism for Enhancing Semanticity and Interpretability of Knowledge Graph Completion Task

In recent years, it has witnessed the revival of knowledge engineering, which is characterized by the rapid growth of Knowledge Graph (KGs). Currently, Knowledge Graph Representation (KGR) and Completion (KGC) methods all use the methodology of scoring triple (
$e_{h},r, e_{t}$
) to judge its correctness. Wherein, in the procedure of conventional Negative Sampling (NS) and reasoning, it is easy to be affected by the low-quality and uncertainty of negative triples related to the given KG, resulting in some unpredictable and inexplicable wrong results that do not conform to commonsense. On the contrary, human beings with commonsense could easily solve this problem. For this reason, this paper proposes a novel mechanism to automatically capture commonsense (concept emphasized here) associated with current KG to provide more deterministic and interpretable semantic information from the fact triple, and further designs a concept-driven negative sampling strategy for commonsense understanding. Overall, a hybrid reasoning approach jointly driven by concept constraints (inter with KG) and fact triple guides (intra with KG), is presented to promote more effective learning and training procedure of KG's representation and completion. Experiments on real-world datasets, show that the proposed work could be used as a flexible plug-in unit to enhance the previous KG representation and completion models.