Joint Probabilistic Reasoning About Coreference and Relations of Univeral Schema
Andrew McCallum, Nicholas Monath · 2017
Abstract : In this project, McCallums IESL lab at UMass Amherst researched and developed technologies for (1) automatic construction of knowledge bases from natural language text corpora, as well as (2) inference on these knowledge bases. Our work proposes and advances Universal Schema, which jointly learns embedded vector representations for the union of all input schema types (relation types, entity types, and entities themselves), including those from existing knowledge bases (such as Freebase and Wikipedia) as well as relations and types in natural language textual patterns. We present techniques for relation and type prediction based on matrix factorization.