Tax assessment is a complex system which is hard to build a accurate mathematical model,furthermore,it is a typical recognition problem of model.to do tax assessment by using probabilistic neural network has its special advantages.the use of PNN algorithm depends greatly on the choosing of sample.whether the chosen sample can reflect the general information characteristics determines the recognition effects of classifying machine.In this article,K-means method is used to get the cluster center of tax payers' information sample to choose sample as the training sample of PNN,thus reaching optimization of PNN algorithm.The research result shows that the classifying effects of improved PNN algorithm is good and has its values of tax assessment.