Bond immunization model is used to minimize interest rate risk for investing in fixed-income market, the model equalizes asset and liability values using the duration which is the sensitivity of portfolio value with respect to the interest rate. However this model might generate an error in practice because the model is based on unreal hypothesis, so called "Parallel Shift Term Structure". In this paper, we use the neural network approach to overcome the parallel shift term structure and try to employ this term structure function to the traditional immunization model. Finally, we present some computational test results that show the superiority of the partial immunization model to the traditional methods.