Institute of Cognitive Science, Universtity of Osnabrück
24/09/2009 à 10h00
Amphithéatre F107, INRIA Grenoble Rhône-Alpes, Montbonnot Saint-Martin
The talk builds around a central finding that concept similarity can be measured on the basis of small sets of characteristic features, or variables, which describe the concepts instances. These variables are selected separately and independently for every concept of two heterogeneous source ontologies populated with text documents by the help of a variable selection procedure of somekind. We define a parameter-dependent similarity measure, which comparessets of characteristic variables for two concepts and we compare its performance with standard parameter-free statistical correlation coeficients. We present empirical results evaluating the performance of the suggested similarity measures by the help of standard variable selection techniquesfor text categorization, methods from descriptive statistics (Discriminant Analysis), as well as a novel variable selection criterion based on binary classification with Support Vector Machines. We provide an evaluation and comparison of the performance of these techniques with respect to the task of inter-ontology concept mapping.
Finally, an overall procedure for ontology matching is built on the suggested above measures of concept similarity, by optimizing the search of similarpairs of concepts, using the structural relation of the concepts within each ontology and recursively applying the similarity criteria on smaller sets of classes. We are lead by the heuristics that similar concepts are found among the descendants of similar concepts in the ontologies hierarchies.