中国科学院自动化研究所机构知识库

Knowledge graph embedding acts as a pivotal role in predicting the missing information in knowledge graphs (KGs). Due to the evolving nature of real-world KGs, one requires the ability to make predictions for newly emerging entities besides those already observed at training time. Current studies have made great efforts to develop a neighborhood aggregator and embed out-of-knowledge-graph (OOKG) entities inductively, with less focus on exploiting the similarity between the existing and newly emerging enti-ties. Attaching importance to such similarity helps facilitate semantic transfer. In this work, we propose a similarity-aware aggregation network for embedding out-of-knowledge-graph entities. Motivated by the fact that similar entities are likely to occur in common graph context, we skillfully design a similarity-aware function, which measures the distance of each entity pair based on the contextual gap. Moreover, we aggregate the neighborhood surrounding the target entity and its similarity information by query-specific attention weights, which are optimized during the learning process. Extensive experi-ments on knowledge graph completion task show that our method achieves substantial improvements over baselines.(c) 2022 Elsevier B.V. All rights reserved.