Knowledge Graph (KG) completion research usually focuses on densely connected\nbenchmark datasets that are not representative of real KGs. We curate two KG\ndatasets that include biomedical and encyclopedic knowledge and use an existing\ncommonsense KG dataset to explore KG completion in the more realistic setting\nwhere dense connectivity is not guaranteed. We develop a deep convolutional\nnetwork that utilizes textual entity representations and demonstrate that our\nmodel outperforms recent KG completion methods in this challenging setting. We\nfind that our model's performance improvements stem primarily from its\nrobustness to sparsity. We then distill the knowledge from the convolutional\nnetwork into a student network that re-ranks promising candidate entities. This\nre-ranking stage leads to further improvements in performance and demonstrates\nthe effectiveness of entity re-ranking for KG completion.\n