Abstract: Knowledge graph completion (KGC) aims at complementing missing entities and relations in a knowledge graph (KG). Popular KGC approaches based on KG embedding are typically limited to the transductive setting, i.e., all entities must be seen during training, which is impractical for real-world KGs where new entities are emerging daily. Recent inductive KG embedding approaches propose to train a neighborhood aggregator in conjunction with entity and relation embeddings, which helps to embed unseen entities via existing neighbors. However, existing methods do not fully take advantage of the structural information of neighbors and are unable to handle triplets involving unseen relations. In this paper, we work further and propose a novel and unified inductive KGC model, namely structure-aware attention network (SAAN), which can efficiently generate embeddings of unseen entities and relations by aggregating neighbors with structure-aware attention weights. Unlike conventional embedding-based attention methods, SAAN can naturally learn importance weights by modeling structural correlations between nodes in an embedding-independent manner and can be applied to any existing KG embedding model. Experimental results on both transductive and inductive KGC tasks show that our model significantly outperforms state-of-the-art methods. Graphical abstract: [Figure not available: see fulltext.]