Adding Distributional Semantics to Knowledge Base Entities through Web-scale Entity Linking

Abstract

Web-scale knowledge bases typically consist entirely of predicates over entities. However, the distributional properties of how those entities appear in text are equally important aspects of knowledge. If noun phrases mapped unambiguously to knowledge base entities, adding this knowledge would simply require counting. The many-to-many relationship between noun phrase mentions and knowledge base entities makes adding distributional knowledge about entities difficult. In this paper, we argue that this information should be explicitly included in web-scale knowledge bases. We propose a generative model that learns these distributional semantics by performing entity linking on the web, and we give some preliminary results that point to its usefulness.