Learning to Generate Ontologies: Using support vector machines to classify ontology triples

Abstract

Ontology engineering is a strenuous task, which has fueled research investigating how an ontology can be generated from text. Ontologies store information in triples, which are also the targets of information extraction. This insight is used in this thesis to motivate using a method from information extraction to learn ontology triples. The transfer of methods is novel, as well as the scope of the triple extraction. Parts of the triple have been extracted with classifiers in previous studies, but in this thesis, a machine learning method is used to classify complete triples. Several classifiers are made to classify triples and arguments of triples. The features used are words, POS tags and dependency parses. The best-performing triple classifier achieves an F1 score of 0.520, outperforming the state of the art, albeit on an easier task. These results indicate that the novel approach taken in this thesis is promising.