For block-oriented Hammerstein systems with a static nonlinear part and a dynamic linear part, there exists a problem of the parameter coupling in nonlinear part and linear part. Traditional methods are to express its output into a linear or a quasi linear regression equation about parameters. However, a Hammerstein system with a neural network (NN) nonlinear part is difficult to be expressed as a linear regression equation about weights and parameters. This paper decomposes parameter coupling by substituting the nonlinear NN equation into the separated key-term of the linear part through the key-term separation idea, and casts the system model into three fictitious models through the hierarchical decomposition principle. Then a hierarchical gradient algorithm is adopted to alternatively identify parameters of these three models. The advantages of the presented Hammerstein system are of the mapping ability of NNs, and the memory ability of dynamic models.