Active Knowledge Graph Completion

Abstract

Knowledge Graphs (KGs) proliferating on the Web are well known to be incomplete. Much research has been proposed for automatic completion, sometimes by rule learning, that is known to scale well. All existing methods learn closed rules. In this paper, we introduce open path (OP) rules and present a novel algorithm, OPRL, for learning OP rules. While CP rules complete a KG by answering given queries, OP rules identify the incompleteness of a KG by generating such queries. For our learning to scale well, we propose a novel, efficient, embedding-based fitness function to estimate the quality of rules. We also introduce a novel, efficient vector computation to formally assess the quality of such rules against a KG. We use adaptations of Freebase, YAGO2, Wikidata, and a synthetic but complete Poker KG to evaluate OPRL. We find that OPRL mines hundreds of accurate rules from massive KGs with up to 8M facts. The learnt OP rules are used to generate queries with precision as high as 98% and recall of 62% on a complete KG, demonstrating the first solution for active knowledge graph completion.