In this paper, a tensor-based semi-blind receiver model is considered for cooperative multiple input multiple output (MIMO) relay systems. The proposed receiver allows a joint symbol and channel estimation without requiring pilot sequences for channel state information (CSI) acquisition. A Tucker-2 tensor model with special structure is constructed. Then we utilize the decomposition uniqueness of the tensor model and design an efficient tensor fitting algorithm. By making full use of the gain of cooperative diversity, the proposed receiver can effectively estimate the information symbols transmitted by the source node and channel matrices even when the number of antennas of the relay and destination nodes is very small. Besides, compared with the existing ALS receiver, the proposed receiver has better signal and channel estimation performance. Simulation results verify the effectiveness of the proposed semi-blind receiver.