The invention belongs to the technical field of natural language treatment, particularly provides a Chinese chapter relation classification method and device and aims at solving the problem of error transfer in a conventional pipeline system method. The Chinese chapter relation classification method comprises the steps of performing distributed representation of a sentence pair on a sentence in aChinese chapter to obtain first sentence pair distributed representation vectors; calculating a similarity and a weight between a memory unit and the first sentence pair distributed representation vectors to obtain memory information of the first sentence pair distributed representation vectors; performing linear combination on the first sentence pair distributed representation vectors and the memory information to generate second sentence pair distributed representation vectors; classifying the second sentence pair distributed representation vectors to obtain a Chinese chapter relation classification result. According to the method, semantic and structural abstraction features inside the sentence are obtained through a deep learning network and the chapter classification effect with superior performance can be obtained.