Toward Automatic Ontology Curation with Similarity-Based Reasoning

Building and maintaining large, reusable ontologies is a prerequisite for building automated systems that reason and plan in a changing world. Unfortunately, our needs, best practices, and mental models of the world also change. These factors— in addition to common human mistakes— lead to model drift over time in our computational models and ontologies. As a result, ontologies contain direct inconsistencies, anomalies, partial or duplicate descriptions, and erroneous constraints between components. This paper presents ongoing development on our system, Marshal, designed to help to curate ontologies and task models by making suggestions while users write procedures and plans. We present experiments in three ontology curation settings: (1) inducing new categories; (2) detecting anomalous instances; and (3) classification of unlabeled instances into an existing ontology. On each task, we compare multiple similarity-based inference strategies, and we show that structure-mapping produces favorable results on all three ontology curation tasks.