Abstract: Efficiently assigning citizen issues in citizen mailboxes to specific departments has become a key factor in improving government service quality and optimize execution efficiency. Deep learning methods usually regard this task as a classification problem and directly give answers in an end-to-end manner. To improve the interpretability, we propose an intelligent assignment algorithm based on retrieval enhancement. Specifically, in the training stage, in addition to the cross-entropy loss of classification, we introduced a supervised contrastive learning loss to enhance the classification and representation capabilities of the model through multi-task training. In the inference stage, we propose a prediction algorithm that combines classification and retrieval, which not only improves accuracy but also makes the results interpretable. Experiments on real data have verified the effectiveness of our proposed method, which can greatly reduce the labor cost and improve efficiency.