Abstract : In this paper, we first introduce two new classes of constrained tensor models that we call generalized PARATUCK- (N1, N) and Tucker- (N1, N) models. A new tensor space-time-frequency (TSTF) coding structure is then proposed for MIMO OFDM-CDMA wireless communication systems. Two semi-blind receivers relying on the new generalized PARATUCK model are derived for solving the problem of joint channel and symbol estimation. One is iterative and based on a two-step alternating least squares (ALS) algorithm. The other one is a closed-form and low-complexity solution which consists of the Kronecker product least squares (KPLS) estimation of the symbol matrix and a matrix unfolding of the channel tensor. Uniqueness of the underlying tensor model is discussed and system design requirements are derived for applicability of the ALS and KPLS receivers. We also show that the so-called TSTF system can be viewed as an extension of three existing tensor-based ST/TST/STF coding systems that are described in a unified framework. Computer simulation results illustrate the good performance of the TSTF system which outperforms the considered existing tensor-based systems both in terms of symbol estimation quality and allocation flexibility.