reduc human overhead text categor arnd christian nig eric brill applic text process requir signific human effort label larg document collect learn statist model extrapol rule knowledg engin work describ reduc effort retain method accuraci construct hybrid classifi util human reason automat discov text pattern complement machin learn standard sentiment classif dataset real custom feedback data demonstr result techniqu result signific reduct human effort requir obtain classif accuraci hybrid text classifi result signific boost accuraci machin learn base classifi compar amount label data 