Abstract: The automatic generation of repetition instruction in air traffic control simulator can improve the intelligence level of the simulator. However, the common problems of polysemous words and synonyms in natural language will affect the quality and effect of the generation task. In order to solve the above problems, a multi-task learning based instruction repetition generation model is proposed. Multi-task learning was introduced in the model, using text instruction comprehension tasks to assist the generation of repetition text and slot filling and intention understanding to constrains the semantic information of vocabulary and sentences. In the training phase, the gradient normalization optimization algorithm was introduced to dynamically update the loss weights of multiple tasks. Experiments were conducted on the air traffic control data set in a real environment. The results show that the accuracy of the proposed model for instruction repetition is significantly improved compared with the baseline model.