build semant kernel text classif wikipedia pu wang carlotta domeniconi document classif present difficult challeng due sparsiti high dimension text data complex semant natur languag tradit document represent word base vector bag word bow dimens term dictionari word corpu simpl commonli represent limit essenti emb semant inform conceptu pattern order enhanc predict capabl classif algorithm paper overcom shortag bow approach embed background knowledg deriv wikipedia semant kernel enrich represent document empir evalu real data set demonstr approach successfulli achiev improv classif accuraci respect bow techniqu recent develop method 