The invention discloses a convolutional neural network-based document automatic question and answer system construction method. The method comprises the following steps of: 1, constructing a theme document library; 2, constructing a word vector model; 3, carrying out theme matching; 4, constructing a word vector matrix; 5, carrying out semantic matching on the basis of a semantic model of a convolutional neural network, wherein the semantic model of the convolutional neural network is divided into three layers, the first layer is a convolutional neural network layer, the second layer is an attention layer and the third layer is a full connection layer; and 6, an answer selection process: selecting a matched answer. According to the method, a synonym dictionary does not need to be manuallyconstructed so that plenty of manpower and time cost are saved, semantic meanings of word contexts can be purposely sampled in a model training process, and an attention mechanism is added in the network so that the contribution degrees, to semantic meanings of whole sentences, of certain representative words can be enhanced.