Leveraging Semantics for Large-Scale Knowledge Graph Evaluation

Knowledge graphs (KG) are being used extensively in different industries for data driven applications. These industrial knowledge graphs, due to their large scale and heterogeneity, are often constructed using automated information extraction (IE) toolkits. Owing to the diverse nature of the sources, such extractions are often noisy and contain many semantic inaccuracies. High quality, consistent KGs are critical to effective predictive analytics and decision support. For example, many commercial question answering systems rely heavily on accurate and consistent knowledge graphs generated from life sciences content. These systems typically require an extensible, scalable, and generalizable framework. To address these issues, we build on previous work in ontology and instance data evaluation and propose a method for Large-Scale Knowledge Graph Evaluation. The approach leverages domain ontologies to detect possible inconsistencies. We construct an RDF/RDFS knowledge graph from the output of a state-of-the-art biomedical IE system, ODIN, and demonstrate that it is easy to construct general inconsistency rules for quality control. In this paper we present our results after applying these rules to the KG and then discuss how our approach and implementation can generalize to many large scale industrial knowledge graphs.