This study proposes a semi‐blind receivers for two‐way multiple‐input multiple‐output (MIMO) relaying systems capable of jointly estimating the channels and symbols in a direct‐data approach. Resorting to a Khatri‐Rao coding scheme applied at both uplink and downlink transmission phases, the authors show that the signals received at the relay and user node are third‐order tensors satisfying the parallel factor (PARAFAC) and PARATUCK2 models, respectively. Combining these two tensor models, a semi‐blind receiver based on an integrated alternating least squares algorithm is proposed to estimate the channels and symbols transmitted by the user nodes without training sequences. The effectiveness of the proposed semi‐blind receiver is corroborated with numerical results, which show that the proposed semi‐blind receiver outperforms state‐of‐the‐art two‐stage training sequence estimator, while operating close to the tensor‐based channel estimator.