ASKRL: An Aligned-Spatial Knowledge Representation Learning Framework for Open-World Knowledge Graph

Ziyu Shang, Peng Wang, Yuzhang Liu, Jiajun Liu, Wenjun Ke

Abstract: Knowledge representation learning (KRL) aims to project entities and relations in knowledge graphs (KGs) to densely distributed embedding space. As the knowledge base expands, we are often presented with zero-shot entities, often with textual descriptions. Although many closed-world KRL methods have been proposed, most of them focus on connections between entities in the existing KGs. Therefore, they cannot handle zero-shot entities well, resulting in the inability of bringing zero-shot entities to existing KGs. To address this issue, this paper proposes ASKRL, a straightforward yet efficient open-world knowledge representation learning framework. ASKRL learns representations of entities and relations in both structured and semantic spaces, and subsequently aligns the semantic space with the structured space. To begin with, ASKRL employs the off-the-shelf KRL models to derive entity and relation embeddings in the structured embedding space. Afterward, a Transformer-based encoder is applied to obtain contextualized representations of existing entities and relations in semantic space. To introduce structure knowledge of KG into the contextualized representations, ASKRL aligns semantic embedding space to structured embedding space from the perspective of common properties (i.e., angle and length). Additionally, it aligns the output distribution of the score function between the two spaces. To further learn representations of zero-shot entities effectively, a sophisticated three-stage optimization strategy is devised in the training phase. In the inference phase, representations of zero-shot entities can be directly derived from the Transformer-based encoder. ASKRL is plug-and-play, enabling off-the-shelf closed-world KRL models to handle the open-world KGs. Extensive experiments demonstrate that ASKRL significantly outperforms strong baselines in open-world datasets, and the results illuminate that ASKRL is simple and efficient in modeling zero-shot entities.