In practical application of undesirable text information identification, most of the text always have intersection even doped with each other. The nonlinear non-separable problem has brought difficulty to undesirable text information identification. SVM can make a nonlinear problem in the original space into a linear problem in high dimension space by nonlinear transformation. And the key of the SVM is to choose the appropriate kernel function. A single kernel function can not recognize the independent undesirable vocabulary and vocabulary combination at the same time, so the recognition accuracy rate is not high and the Rcall value is not ideal. For the specific application of undesirable text information identification, combining with linear kernel and homogeneous polynomial kernel it structured a new combination kernel function according to the Mercer theorem. This combination kernel function has the advantage of both linear kernel and polynomial kernel, and could identify the independent undesirable vocabulary and vocabulary combination. Then it evaluated the linear kernel, homogeneous polynomial kernel and combination kernel function in the sample experiment. The experimental results showed that the recognition accuracy rate and the Rcall value of combination kernel function was more ideal than other kernel functions.