Abstract

The task of knowledge graph construction presents a confounding challenge for statistical relational models. While the uncertainty of extractions from NLP tools and the ontological structure of knowledge are a perfect match for the strengths of statistical relational techniques, the vast and continually growing evidence from which knowledge graphs are constructed can make such models prohibitively expensive. We address this challenge by presenting two lines of research that provide a foundation for online knowledge graph construction. The first is work on knowledge graph identification, a scalable probabilistic model for combining statistical features from uncertain extractions and ontological constraints to efficiently construct a knowledge graph. The second is the necessary theory and accompanying algorithms for partially updating an inferred knowledge graph. We illustrate how combining these components presents the opportunity to apply sophisticated statistical relational models to complex domains, such as knowledge graph construction, without sacrificing quality or efficiency.

Date

March 4, 2026

Authors

Jay Pujara, Ben London, Lise Getoor, William W Cohen

Journal

Fifth International Workshop on Statistical Relational AI