Lightweight ontology learning using dependency parsing

Abstract

An ontology is a knowledge representation structure composed of concepts which are linked together through relations. Ontologies strike the right balance between a formally defined structure and a simple easy to read artefact. Due to this balance, ontologies are both human and machine readable. Ontologies can be used as a machine readable, semantical representation of natural language documents. The vision of the Semantic Web has exactly this idea of using ontologies to encode semantics on the web, as its core ambition. The demand for ontologies has been growing in recent years, and as a result the field of Ontology Learning was born. Ontology learning refers to a set of techniques which help ontologists in the process of constructing an ontology. These techniques try to automate as much as possible the whole process of knowledge acquisition and modelling, through the use of machine learning, statistical approaches and Natural Language Processing techniques amongst others. The system presented in this thesis is a semi-automatic lightweight ontology learning system which operates on unstructured English text. The system is based on the assumption that there exists a correlation between grammatical relations in sentences and semantic relations between concepts. Through a set of binary classifiers (Support Vector Machines), the application attempts to uncover the relationship between grammatical relations and semantic relations. Every classifier is specifically build to predict a particular semantic relation (example is-a) between noun phrases in a sentence. Sentences are encoded as dependency parse graphs, and semantic relations between noun phrases in the sentence are represented by the shortest graph path between their corresponding vertices. The extracted relations are transformed into and ontology by applying a simple mapping. Two classifiers for the is-a and the part-of relations have been built, and they extract relations from text with an F-Measure ranging from 45% till 50%.