NASTyLinker: NIL-Aware Scalable Transformer-Based Entity Linker

Entity Linking (EL) is the task of detecting mentions entities in text and disambiguating them to a reference knowledge base. Most prevalent EL approaches assume that base complete. In practice, however, it necessary deal with case linking an entity not contained (NIL entity). Recent works have shown that, instead focusing only on affinities between entities, considering inter-mention can be used represent NIL by producing clusters mentions. At same time, help substantially improve performance for known entities. With NASTyLinker, we introduce approach aware produces corresponding mention while maintaining high The based dense representations from Transformers resolves conflicts (if more than one assigned cluster) computing transitive mention-entity affinities. We show effectiveness scalability NASTyLinker NILK, dataset explicitly constructed evaluate respect Further, apply presented actual task, namely graph population Wikipedia listings, provide analysis outcome.