gener compon analysi text heterogen attribut xuerui wang chri pal andrew mccallum present class richli structur undirect hidden variabl model suitabl simultan model text attribut encod modal model gener techniqu princip compon analysi heterogen data type contrast approach framework modal word author timestamp captur natur probabilist encod latent space represent previous unseen document obtain fast matrix multipl method demonstr effect framework task author predict year nip confer proceed recipi predict task month academ email archiv research approach broadli applic real world applic wish effici make predict larg number potenti output dimension reduct defin probabilist framework 