The invention relates to a multi-level natural language anti-junk text method and system. The method comprises the steps of obtaining a sensitive word recognition result and a sensitive word deformation recognition result of a to-be-recognized text; and performing text classification on the to-be-identified text to obtain a pre-judgment probability that the to-be-identified text is a junk text, and performing comprehensive judgment based on the sensitive word identification result, the sensitive word deformation identification result and the pre-judgment probability to obtain a final probability that the to-be-identified text is the junk text. According to the method, the junk text can be efficiently recognized, the adverse effect of the junk text on the Internet health communication environment can be avoided, the robustness is higher, and the method can be widely applied to the social contact, comments and other Internet products.