The Role of Knowledge in Determining Identity of Long-Tail Entities

Abstract

Identifying entities in text is an important step of semantic analysis. Some entity mentions comprise a name or description, but many include no information that identifies them in the system’s knowledge resources, which means that their identity cannot be established through traditional disambiguation. Consequently, such NIL (not in lexicon) entities have received little attention in entity linking systems and tasks so far. However, given the non-redundancy of knowledge on NIL entities, their lack of frequency priors, their potentially extreme ambiguity, and their numerousness, they constitute an important class of long-tail entities and pose a great challenge for state-of-the-art systems. In this paper, we describe a method for imputing identifying knowledge to NILs from generalized characteristics. We enrich the locally extracted information with profile models that rely on background knowledge in Wikidata. We describe and implement two profiling machines using state-of-the-art neural models. We evaluate their intrinsic behavior and their impact on the task of determining the identity of NIL entities.

Keywords: Long-tail entities, NIL clustering, knowledge-based completion