Extracting Domain Ontologies with CORDER

The CORDER web mining engine developed at the Knowledge Media Institute computes a lexical coocurrence network out of websites - a binary relation R. A natural extension of CORDER would be that of learning an ontology. However, our work shows that coocurrence proves insufficient to discover concepts and conceptual taxonomies (i.e. very simple ontologies) out of this network. To tackle this problem two unsupervised learning methods were studied based, on the one hand, on set similarity (and thus on a set-based representation of the data) and, on the other hand, on cosine similarity (and thus on a vector-space representation of the data). The underlying idea being that of taking into account, for the clustering, as features, their related coocurring entities (and thus the indirect links among the entities), as suggested, for instance, by O. Ferret. For the purposes of this study, we restricted ourselves to (solely) research areas. The most promising results in our experiments were given by the vector-space representation. To validate the results we used the ACM classification of computer science research areas as our gold standard.