We study the problem of embedding-based entity alignment (EA) between knowledge graphs (KGs), and propose a novel framework that uni es multiple views of entities to learn their embeddings. Experiments on real-world datasets show that this framework largely outperforms the current embedding-based methods. Entity alignment (EA) aims to nd entities in di erent knowledge graphs (KGs) referring to the same real-world identity. Conventional methods identify similar entities based on the symbolic features, such as names, textual descriptions and attribute values. Recently, increasing attention has been drawn to leveraging the KG embedding techniques for dealing with this problem, where the key idea is to learn vector representations (called embeddings) of KGs and nd alignment according to the similarity of the embeddings. We propose a new EA framework, MultiKE, based on multi-view KG embedding. The underlying idea is to divide the various features of KGs into multiple subsets (called views), which are complementary to each other (see Figure 1 for example). Thus, entity embeddings can be learnt from each separate view and jointly optimized to improve the alignment performance.
Multi-view KG embedding. Based on the data model of KGs, we de ne three representative views based on the name, relation and attribute features. First, literals are constituted by sequences of tokens. We embed the name view using the literal embeddings. Second, the relation view characterizes the structure of KGs. We employ TransE to interpret a relation as a translation vector from its head entity to tail entity. Third, for the attribute view, we use a convolutional neural network to extract features from the attributes and values of entities. Cross-KG training. We propose the cross-KG entity identity inference to capture the alignment information using seed alignment. We also present the crossKG relation/attribute identity inference to enhance EA. ? Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Mona Lisa Leonardo da Vinci – other title : La Gioconda author – description : Italian Renaissance polymath – year : 1503 – birth name : Leonardo di ser Piero da Vinci … … Louvre – caption : The Richelieu Wing – established : 1793 …
birth place
Attributes JAPE [
MultiKE-WVA MultiKE-SSL MultiKE-ITC
DBP-WD DBP-YG Hits@1 Hits@10 MR MRR Hits@1 Hits@10 MR MRR 31.84 58.88 266 0.411 23.57 48.41 189 0.320 57.19 69.53 182 0.618 42.71 48.30 137 0.446 38.96 66.77 142 0.487 23.24 42.70 706 0.300 90.42 94.59 22 0.921 85.92 94.99 19 0.891 91.86 96.26 39 0.935 82.35 93.30 21 0.862 91.45 95.19 114 0.928 88.03 95.32 35 0.906 View combinations. Intuitively, general entity embeddings can bene t from multiple view-speci c embeddings. We propose weighted view averaging (WVA), shared space learning (SSL) and in-training combination (ITC). 3
We selected two datasets in [