The invention discloses a junk short message identification method combining word vectors with machine learning. The method comprises the following operation steps of (1) carrying out the first-step identification on a junk short message according to the short message characteristics; (2) carrying out the second-step identification on the junk short message according to the keywords; (3) calculating a short message text vector of the short message, and performing the third-step recognition on the junk short message by using a method of a support vector machine; (4) calculating a static word vector matrix of the short message, and performing the fourth-step recognition on the iunk short message by using a convolutional neural network; and (5) calculating a dynamic word vector of each segmented word of the short message, and carrying out the fifth-step recognition on the junk short message by using the convolutional neural network. According to the method, an unsupervised and supervisedcombined junk short message identification method is adopted, so that the identification accuracy of the junk short messages can be greatly improved.