A structure-enhanced generative adversarial network for knowledge graph zero-shot relational learning

Abstract

Most knowledge graph completion methods focus on predicting existing relationships in the knowledge graph but cannot predict unseen relationships. To solve this problem, knowledge graph zero-shot relational learning (KGZSL) has gotten more and more attention in recent years. The common method of KGZSL is to leverage Generative Adversarial Networks (GANs) to build the connection between existing relation descriptions and knowledge graph domains. However, the traditional KGZSL method ignores the gap between relation text description and relation structured representation. To bridge this gap, we propose a Structure-Enhanced Generative Adversarial Network (SEGAN). SEGAN adopts a structure encoder to introduce knowledge graph structure information into the generator and guide the generator to generate knowledge graph embeddings more accurately. In addition, in the KGZSL task, the representations of entities are closely tied to the existing relationships, which has a negative impact on the prediction of new instances. Therefore, we design a new plug-and-play feature encoder to decouple entities from existing relationships. Experimental results on the knowledge graph zero-shot relational learning dataset demonstrate that our method has better structure representation ability, and the model performance is improved by 36.3% compared with the current optimal model.