Exception-Enriched Rule Learning from Knowledge Graphs.

Advances in information extraction have enabled the automatic construction of large knowledge graphs (KGs) like DBpedia, Freebase, YAGO and Wikidata. These KGs are inevitably bound to be incomplete. To fill in the gaps, data correlations in the KG can be analyzed to infer Horn rules and to predict new facts. However, Horn rules do not take into account possible exceptions, so that predicting facts via such rules introduces errors. To overcome this problem, we present a method for effective revision of learned Horn rules by adding exceptions (i.e., negated atoms) into their bodies. This way errors are largely reduced. We apply our method to discover rules with exceptions from real-world KGs. Our experimental results demonstrate the effectiveness of the developed method and the improvements in accuracy for KG completion by rule-based fact prediction.