Towards SHACL learning from knowledge graphs

Abstract

Knowledge Graphs (KGs) are typically large data-first knowl- edge bases with weak inference rules and weakly-constraining data schemes. The SHACL Shapes Constraint Language is a W3C recommendation for the expression of shapes as constraints on graph data. SHACL shapes serve to validate a KG and can give informative insight into the structure of data. Here, we introduce Inverse Open Path (IOP) rules, a logical for- malism which acts as a building block for a restricted fragment of SHACL that can be used for schema-driven structural knowledge graph validation and completion. We define quality measures for IOP rules and propose a novel method to learn them, SHACLearner. SHACLearner adapts a state-of-the-art embedding-based open path rule learner (oprl) by modifying the efficient matrix-based evaluation module. We demonstrate SHACLearner performance on real-world massive KGs, YAGO2s (4M facts), DBpedia 3.8 (11M facts), and Wikidata (8M facts), finding that it can efficiently learn hundreds of high-quality rules.