Multi-category Visual Entity Linking Using Wikimedia

Abstract

This research creates a dataset for visual entity linking. Visual entity linking is the task that links objects in an image to an entity in a knowledge base. Visual entity linking makes it possible to use more than just visual information using a knowledge base. However, existing datasets related to visual entity linking have problems with linked entities being at the category level, limited categories, small number of images per entity, and entities being ambiguous. We create a dataset consisting of an entity name, Wikidata ID, and images corresponding to each entity across multiple categories, from Wikimedia Commons. We analyze the dataset and train and evaluate an image recognition model on the created dataset for visual entity linking.