Retrosynthesis prediction of chemical reactions is a fundamental yet challenging problem in the field of chemistry. Existing methods, especially those based on deep learning such as Transformer models, tend to learn the map between reactants and products directly from chemical reaction datasets. However, these models often neglect prior knowledge from the chemistry domain. Without the prior knowledge, the performance of models heavily relies on the quality of reaction dataset, instead of understanding molecule. To address these issues, we propose the SMART-RetroNet framework. Our approach involves a pre-training task of mapping different SMILES representations to their canonical forms, during which the model could successfully learn fundamental prior knowledge in chemistry field. Experiment results demonstrate that SMART-RetroNet can achieve significant performance improvements in chemical reaction one-step retrosynthesis task. This work provides an effective approach for chemical reaction retrosynthesis prediction and a way to transfer prior knowledge in related chemistry tasks.