Entity linking with a unified semantic representation

Abstract

Entity Linking (EL) consists in linking mentions in a document to their referent entities in a Knowledge Base. Current approaches fall into two main categories: local approaches, in which mentions are linked independently of each other, and global approaches, in which all mentions are linked collectively. Local approaches often ignore the semantic relatedness of entities, and while global approaches incorporate the semantic relatedness, they tend to focus only on directly connected entities, ignoring indirect connections which might be useful. We present a global EL approach that unifies the representation of the semantics of entities and documents--the probability distribution of entities being visited during a random walk on an entity graph--that accounts for direct and indirect connections. An experimental evaluation shows that our method outperforms five state-of-the-art EL systems and two very strong baselines.