fast logist regress text categor variabl length gram georgiana ifrim khan bakir gerhard weikum common represent text categor bag word model aka unigram model learn represent involv typic preprocess stopword remov stem result explicit token corpu work introduc logist regress approach learn involv automat token weaken priori requir knowledg corpu result token variabl length word charact gram basic token accomplish solv logist regress gradient ascent space ngram show effici branch bound approach choos maximum gradient ascent direct project singl dimens candid featur space larg method investig variabl length gram learn demonstr effici approach compar state art classifi text categor cyclic coordin descent logist regress support vector machin 