The term structure of interest rates is a basic problem in financial field. Especially in the process of Chinese marketization of interest rates, research on the term structure of interest rates has very important theoretical and practical significance to the development and improvement of Chinese financial market. In this paper, we take advantage of faster learning speed, stronger capability of adaptability and numerical approximation of neural network characteristics to make the empirical analysis on the 14 group data selected from the Shanghai Security Exchange Market of Government Bonds traded on 12-Feb-2010 by means of BP and RBF neural network respectively. The results show that neural network has higher accuracy in predicting yields of government bonds, and calibration of parameters can affect the accuracy of network to some extent.