Date of Original Version
Abstract or Description
The contributions of automatic semantic concept classifiers for interactive filtering (classifiers in conjunction with query rankings) and browsing (classifiers in lieu of query rankings) are tested against three test corpora: an amateur photo collection, documentary video, and news video. Results show that current classifiers offer browsing utility twice as good as having no classifier at all, and that continuous improvements in the classifiers produce comparable improvements in the browsing utility. For filtering a wellordered set of results (e.g., a set retrieved from text search), concept classifiers need greater accuracy: current classifiers showed worse performance than not filtering at all, even when the classifiers’ accuracy is nearly doubled. Results are consistent for all test corpora. Hence, automatic semantic concepts can offer significant utility for browsing at current levels of accuracy, but the requirement is much higher for filtering a well-ordered set of results, where extreme accuracy is necessary before benefits are seen.