This paper presents a new scoring method for the translation in CET-4 and CET-6 based on improved PSO-BP neural network. We get the text feature values by using the algorithm of BLEU and SVD, together with the scores the teacher have scored, are gathered as input set. We use it to train the improved PSO-BP neural network, which can be reversely used to predict the translation score. This paper improves the PSO-BP neural network by the calculation of the inertia weight and the adaptive value function. We use Matlab to make simulation, the result shows that, in the translation scoring, the use of improved PSO-BP neural network is better than using multiple linear regression to obtain better correlation, and the Pearson correlation coefficient of artificial scoring average increased by 12%.