Classifying Candidate Axioms via Dimensionality Reduction Techniques

We assess the role of similarity measures and learning methods in classifying candidate axioms for automated schema induction through kernel-based learning algorithms. The evaluation is based on (i) three different similarity measures between axioms, and (ii) two alternative dimensionality reduction techniques to check the extent to which the considered similarities allow to separate true axioms from false axioms. The result of the dimensionality reduction process is subsequently fed to several learning algorithms, comparing the accuracy of all combinations of similarity, dimensionality reduction technique, and classification method. As a result, it is observed that it is not necessary to use sophisticated semantics-based similarity measures to obtain accurate predictions , and furthermore that classification performance only marginally depends on the choice of the learning method. Our results open the way to implementing efficient surrogate models for axiom scoring to speed up ontology learning and schema induction methods.