Dual Channel Knowledge Graph Embedding with Ontology Guided Data Augmentation

Tengwei Song, Long Yin, Xudong Ma, Jie Luo

Abstract: Current knowledge graph completion suffers from two major issues: data sparsity and false negatives. To address these challenges, we propose an ontology-guided joint embedding framework that utilizes dual data augmentation channels and a joint loss function to learn embeddings of knowledge graphs. Our approach spontaneously generates positive and negative instances from two distinct ontology axiom sets, leading to improved completion rates for originally sparse knowledge graphs while also producing true-negative samples. Additionally, we propose two novel metrics for evaluating a model’s reasoning capabilities in predicting relations or links using KG and ontology data, thus avoiding incorrect predictions. Empirical results demonstrate that our framework outperforms existing models in most tasks and datasets, with significantly better performance in many cases for reasoning capability evaluation metrics.