Learning Disjointness Axioms Technical Report

Our approach to the automatic acquisition of disjointness axioms relies on a machine learning classifier that determines disjointness of any two classes. The classifier is trained based on a “Gold Standard” of manually created disjointness axioms, i.e. pairs of classes each of which is associated with a label – “disjoint” or “not disjoint” – and a vector of feature values. As in our earlier experiments [7], we used a variety of lexical and logical features, which we believe to provide a solid basis for learning disjointness. These features are used to build an overall classification model on whose basis the classifier can predict disjointness for previously unseen pairs of classes. We implemented all features and auxiliary methods for training and classification within the open-source framework LeDA1 (Learning Disjointness Axioms), a complete redesign and re-implementation of our original prototype. LeDA is open-source and publicly available under the LGPL license.