Per John Quigley's suggestion, Folksonomy takes a tag system, that is a bunch of tags, and learns a relationship between project descriptions and the tags, so that you can tag arbitrary projects based on just their description. So far, it just uses the Debian debtags tag system, but it is trivial to implement for Sourceforge as well as Freshmeat tags. The dataset is the datamart from the Flossmole project, plus the output of an `apt-cache search -f .`
./query-tool.pl tagsearch . ./query-tool.pl search "works-with::image:vector"The program needs to be cleaned up and given to the Debtags people. The models take a long time to learn and use the AI::Categorizer::Learner::SVM perl module.