面向临床发现的复杂事件抽取

A Clinical-Discovery-Oriented Complex Event Extraction

临床发现事件抽取即从电子病历中检测和提取所需事件的属性。由于事件属性的多样性、多事件属性的重叠性、垂直领域语料的专业性、样本分布的不均衡性增加了事件抽取任务的复杂程度,常规的方法无法很好地解决问题。为了适应任务的复杂性,该文提出了一种面向临床发现的管道式事件抽取方法,将事件抽取划分为基于序列标注的触发词抽取、基于指针网络的论元抽取和基于匹配的事件极性预测三个模块。该方法在中国健康信息处理会议(CHIP2021)评测2数据集上获得0.4303的F 1值,取得了第1名的成绩。

Clinical discovery oriented event extraction is to detect and extract the attributes of the required events from the electronic medical records.This task is challenging due to the event attributes diversity,the overlapping of multi-event attributes,the speciality of the domain corpus and samples imbalance,and the conventional method cannot solve the problem properly.This paper proposes a clinical discovery oriented event extraction method including three modules:a trigger extraction module based on sequence labeling,an argument extraction module based on pointer network,and a polarity prediction module based on matching.Tested in CHIP 2021 Track 2"Evaluation of Chinese Clinical Discovery Event Extraction",the method achieves 0.4303 F 1-score as the top score in this competition.