A Hierarchical Knowledge Graph Embedding Framework for Link Prediction

Knowledge graph embedding maps the semantics of entities and relations to a low-dimensional space by optimizing the vector distance between positive and negative triples. Traditional negative sampling techniques usually regard high-scoring triples as high-quality negative triples, but this not only easily introduces false negative triples, but also ignores important information in the graph structure. To address these issues, we propose an easily pluggable hierarchical knowledge graph embedding framework. High-quality corrupted entities are generated through semantic and structural information, and then margin estimation is used to generate high-quality negative triples, and the structural information of the entities is combined to perform link prediction on new facts. Experimental results show that our framework improves the performance of the original knowledge graph embedding model, in which the hierarchical subgraph negative sampling module outperforms other negative sampling techniques. The framework we proposed can be easily adapted to various knowledge graph embedding models and explain the prediction results.