Path-Based Learning for Plant Domain Knowledge Graph

Cuicui Dong, Huifang Du, Yaru Du, Ying Chen, Wenzhe Li, Ming Zhao

Abstract: Learning to embed the knowledge graph has been a hot topic in research communities. As for that, TransE is a promising method that can achieve state-of-art performance for many of the benchmark tasks. However, none of the previous work considers the knowledge graph in plant domain in which case the properties of the graph are significantly different. For the knowledge graph in plant domain, most of its relations belong to one-to-many, many-to-one or many-to-many types (actually majority of them are attribute-type relations), which are not in the scope of consideration for classical TransE model. In order to deal with such unique challenges, we propose a novel model called PTA (path-based TransE for attributes). It constructs the relation path by combining attributes and hyponymy relations, and embeds them to a lower dimensional space as well. We conduct extensive experiments on link prediction task where the performance is measured by mean rank and Hit@10. The results show that our new model significantly outperforms other competing methods on several different tasks.