Learning the distance metric in a personal ontology

Personal ontology construction is the task of sorting through relevant materials, identifying the main topics and concepts, and organizing them to suit personal needs. Automatic construction of personal ontologies is difficult in part because measuring the semantic distance between two concepts is difficult. Knowledge-based approaches use either knowledge bases, such as WordNet, or lexico-syntactic patterns to induce the differences between concepts. However, these techniques are only applicable for a subset of concepts and leave the majority unmeasurable. On the other hand, statistical approaches are able to induce the differences between any concept pair but lack of human knowledge involvement and hence suffer from low precision. In the context of personal ontology construction, semantic distances between concepts need to reflect personal preferences. Based on that, this paper presents a supervised hierarchical clustering framework to incorporate personal preferences for distance metric learning in personal ontology construction. In this framework, periodic manual guidance provides training data for learning a distance metric and the learned metric is used during automatic activities to further construct the ontology. A detailed user study demonstrates that the approach is effective and accelerates the construction of personal ontologies.