In natural language processing, the selection of the token is a very important step. The original text should segment into some granularity, then subsequent processing and analysis work can be carried out. The influence of select different segmentation granularity on the Korean text classification task is discussed. Due to the Korean language composition characteristics, combined with the linguistic knowledge the text was divided into six different levels: phoneme, syllable, subword, word, space writing without suffix, and space writing. Then build text semantic expressions in the vector space model. According to the different granularity performance in five classic classifiers and six deep learning models, Korean text feature representations were analyzed and compared in six different granularities. The Korean scientific literature was classified, and the results show that the spacing without suffix level performs best in seven classifiers, and the highest classification accuracy rate reaches 91.94%.