Optimizing Ontology Learning Systems that Use Heterogeneous Sources of Evidence

As the manual construction of ontologies is expensive, many systems to (semi-) automatically generate ontologies from data have been built. More recently, such systems typically integrate multiple and heterogeneous evidence sources. In this paper, we propose a method to optimize ontology learning frameworks by finding near-optimal input weights for the individual evidence sources. The optimization process applies a so-called source impact vector and the Tabu-search heuristic to improve system accuracy. An evaluation in two domains shows that optimization provides gains in accuracy of around 10 %.