Abstract

The growing popularity of structured knowledge bases such as knowledge graphs necessitates integrating multiple knowledge sources. A key component of this integration is entity resolution (ER), reconciling instances of a single entity occurring in different knowledge graphs. In contrast to the conventional ER problem setting, we consider the scenario where ER judgments for related entities are made collectively while also determining when a new entity should be added to the graph. Our approach uses hinge-loss Markov random fields to define a joint probability distribution over entity coreferences. We apply this model to two publicly-available knowledge graphs, MusicBrainz and Freebase where relational structure allows us to collectively resolve musical artists and albums, achieving an F1 of 0.84.

Date

March 20, 2026

Authors

Jay Pujara, Kevin Murphy, Xin Luna Dong, Curtis Janssen

Journal

Bay Area Machine Learning Symposium