RDF Knowledge Base Summarization by Inducing First-Order Horn Rules

Abstract

RDF knowledge base summarization produces a compact and faithful abstraction for entities, relations, and ontologies. The summary is critical to a wide range of knowledge-based applications, such as query answering and KB indexing. The patterns of graph structure and/or association are commonly employed to summarize and reduce the number of triples. However, knowledge coverage is low in state-of-the-art techniques due to limited expressiveness of patterns, where variables are under-explored to capture matched arguments in relations. This paper proposes a novel summarization technique based on first-order logic rules where quantified variables are extensively taken into account. We formalize this new summarization problem to illustrate how the rules are used to replace triples. The top-down rule mining is also improved to maximize the reusability of cached results. Qualitative and quantitative analyses are comprehensively done by comparing our technique against state-of-the-art tools, with showing that our approach outperforms the rivals in conciseness, completeness, and performance.