关键词: 关系抽取, 联合抽取, 对抗学习, 多特征融合, 关系重叠

Abstract: The entity relationship extraction of Chinese electronic medical records is a key part for constructing medical knowledge graphs and serving downstream tasks.Due to the complex relations in medical texts and high density of entities，inaccurate identification of medical terms and other problems may occur.To address these issues，a model called Adversarial Learning and Multi-Feature Fusion for Relation Triple Extraction-AMFRel is proposed in this paper.The model first extracts texts and part-of-speech features from medical text to obtain encoded vectors that incorporate part-of-speech information.Then，encoding vector is employed to generate adversarial samples by combining the perturbations generated by adversarial training to extract the subject of the sentence.Finally，the model enriches the structural features of the text by using an information fusion module，extracts the corresponding object based on specific relationship information，and obtains a triplet of medical text.Experiments are conducted on the CHIP2020 relation extraction dataset and the diabetes dataset.Our results show AMFRel achieves a precision of 63.922%，recall of 57.279%，and F1 score of 60.418% on the CHIP2020 relation extraction dataset，and a precision of 83.914%，recall of 67.021%，and F1 score of 74.522% on the diabetes dataset，demonstrating the triple extraction performance of this model is superior to other baseline models.