Disentangled Ontology Embedding for Zero-shot Learning
Yuxia Geng, Jiaoyan Chen, Wen Zhang, Yajing Xu, Zhuo Chen, Jeff Z. Pan, Yu‐Feng Huang, Feiyu Xiong, Hua‐Jun Chen
Abstract
Knowledge Graph (KG) and its variant of ontology have been widely used for\nknowledge representation, and have shown to be quite effective in augmenting\nZero-shot Learning (ZSL). However, existing ZSL methods that utilize KGs all\nneglect the intrinsic complexity of inter-class relationships represented in\nKGs. One typical feature is that a class is often related to other classes in\ndifferent semantic aspects. In this paper, we focus on ontologies for\naugmenting ZSL, and propose to learn disentangled ontology embeddings guided by\nontology properties to capture and utilize more fine-grained class\nrelationships in different aspects. We also contribute a new ZSL framework\nnamed DOZSL, which contains two new ZSL solutions based on generative models\nand graph propagation models, respectively, for effectively utilizing the\ndisentangled ontology embeddings. Extensive evaluations have been conducted on\nfive benchmarks across zero-shot image classification (ZS-IMGC) and zero-shot\nKG completion (ZS-KGC). DOZSL often achieves better performance than the\nstate-of-the-art, and its components have been verified by ablation studies and\ncase studies. Our codes and datasets are available at\nhttps://github.com/zjukg/DOZSL.\n