The scale of datasets and knowledge bases is relatively limited in the field of Traditional Chinese Medicine (TCM). To efficiently construct and supplement these knowledge repositories while expanding the scale of datasets, this study proposes an optimized UniLM natural language processing model. Initially, we developed a UniLM model based on transfer learning. Secondly, During the training process, we introduced minor perturbations into the embedding layer to generate adversarial samples, thereby enhancing the model’s generalization capabilities through adversarial training. Finally, we employed a five-fold cross-validation method to validate and evaluate the model. The model achieved an average recall of 0.977, an average accuracy of 0.973, and an average F1 score of 0.975. The optimized UniLM model exhibits high robustness and strong generalization capabilities. These attributes contribute to improved performance in generating questions in the TCM field as well as in the comprehensive utilization of textual data.