An automatic text-classification algorithm based on sequence is presented in this paper. It utilizes the semantic relevance on two levels: relevance between sentences (subpattern) and between keywords which represent specific meaning (concept node) in one sentence. In this way, each keyword can be combined with dynamic weight. For subpatterns which contain no keywords, Markov model is used to estimate the amplitude of their signals, thereby the feature sequence for the text which needs to be classified is created.In the experiment of classifying Chinese documents,it is BEP value is about 83％.Furthermore,it is easy to implement in actual system.