Visual Pivoting for (Unsupervised) Entity Alignment

This work studies the use of visual semantic representations to align entities in heterogeneous knowledge graphs (KGs). Images are natural components many existing KGs. By combining with other auxiliary information, we show that proposed new approach, EVA, creates a holistic entity representation provides strong signals for cross-graph alignment. Besides, previous alignment methods require human labelled seed alignment, restricting availability. EVA completely unsupervised solution by leveraging similarity create an initial dictionary (visual pivots). Experiments on benchmark data sets DBP15k and DWY15k offers state-of-the-art performance both monolingual cross-lingual tasks. Furthermore, discover images particularly useful long-tail KG entities, which inherently lack structural contexts necessary capturing correspondences. Code release: https://github.com/cambridgeltl/eva; project page: http://cogcomp.org/page/publication view/927.