Abstract: 【Purpose/significance】Based on the comparative analysis of existing relation extraction methods, we construct a relation ex⁃ traction model suitable for medical texts.【Method/process】We construct an AGP model to implement relation extraction. This model takes the embedding representation of medical text as input, which is then processed through a transformer encoder for further extrac⁃ tion. A global pointer network is utilized for decoding. To enhance robustness, adversarial training is integrated into the model【. Result/ conclusion】The AGP model achieves F1 scores of 0.6190, 0.5321, and 0.5684 on the CMeIE-V1, CMeIE-V2, and DiaKG datasets, respectively. The experiment results demonstrate that the AGP model outperforms baseline models on medical text relation extraction tasks【. Innovation/limitation】The model proposed in this paper does not integrate large language models.