Knowledge Inference and Knowledge Completion Methods using Neuro-Symbolic Inductive Rules

In knowledge graph completion, a symbolic reasoning method establishes a human readable rule by analyzing an imperfect knowledge graph and infers knowledge omitted by an inference engine. However, the entire rules cannot be defined based on a large-scale knowledge graph. This study proposes a method, based on a knowledge graph, that can facilitate end-to-end learning and induce rules without several processing steps that require direct human involvement. The proposed method combines the concept of unification used in symbolic reasoning and deep learning for training vectors expressing symbols. It trains the vectors expressing relations of rule schemas defined to induce rules based on a given knowledge graph. Furthermore, the performance of the proposed method is evaluated against neural theorem prover and the greedy neural theorem prover, which are recently developed neuro-symbolic models, based on four benchmark datasets. The experimental results verify that the proposed method induces more significant rules in less training time. Furthermore, this study conducted an experiment on knowledge graph completion, implemented by an inference engine. Based on the experiment results, it was confirmed that the rules induced by the proposed model can indeed effectively complete missing knowledge.