Feature selection is a very important process in text classification. It can effectively eliminate redundant features and retain feature words with strong class distinguishing ability. In this paper, we propose a feature selection algorithm based on document frequency of segmented term frequency (STF-DF). In the algorithm, we also present two new concepts of ``segmented term frequency”and ``STF-DF.”Then, we compare STF-DF with six commonly used feature selection algorithms (document frequency, information gain, chi-square, CMFS, NDM, and t-test) on three popular datasets (20 Newsgroups, Classic3, and WebKB). Experimental results show that our proposed algorithm can improve the accuracy of text classification and make the classification more effective.