面向中医医案的事件抽取方法研究  

Research on event extraction methods for medical records of traditional Chinese medicine

作　　者：马月坤 崔漠晓 MA Yuekun;CUI Moxiao(College of Artificial Intelligence,North China University of Science and Technology,Tangshan,Hebei 063210,China;Hebei Provincial Key Laboratory of Industrial Intelligent Perception,Tangshan,Hebei 063210,China;School of Computer and Communication Engineering,University of Science and Technology Beijing,Beijing 100083,China;Beijing Key Laboratory of Knowledge Engineering for Materials Science,Beijing 100083,China)

机构地区：[1]华北理工大学人工智能学院,河北唐山063210 [2]河北省工业智能感知重点实验室,河北唐山063210 [3]北京科技大学计算机与通信工程学院,北京100083 [4]材料领域知识工程北京市重点实验室,北京100083

出　　处：《河北科技大学学报》2025年第2期141-150,共10页Journal of Hebei University of Science and Technology

基　　金：国家重点研发计划项目(2022YFC3502303);河北省工业智能感知重点实验室项目(SZX2021013)。

摘　　要：针对中医事件抽取过程中存在事件论元边界模糊的问题,综合运用卷积神经网络、双向长短期记忆网络和注意力机制提出了一种局部与全局语义特征融合的事件抽取模型(event extraction model integrating local and global semantic feature,EE-LGSF),以提升中医事件抽取的效果。首先,通过结合不同滤波窗口大小的卷积神经网络提取文本的多维度局部特征信息,同时利用双向长短期记忆网络捕捉文本的全局特征信息;其次,通过门控机制实现局部与全局信息的动态交互,以增强对论元边界的识别能力;再次,引入模糊跨度注意力机制,动态调整注意力范围,从而优化论元跨度的决策过程;最后,以条件随机场进行标签预测。结果表明,所提模型比相关模型在中医医案数据集上的F1值提升了3.0~11.0个百分点,在解决中医事件抽取问题方面表现更为优异。所提模型能有效利用文本局部和全局语义信息,提高对文本跨度学习的灵活性来增强模型对论元边界的识别能力,从而获得更好的中医事件抽取效果,对中医知识传承和发展具有参考价值。In response to the issue of fuzzy event argument boundaries in traditional Chinese medicine(TCM)event extraction,an event extraction model integrating local and global semantic features(EE-LGSF)was proposed,which combined convolutional neural networks,bidirectional long short-term memory networks,and attention mechanisms to enhance the effectiveness of TCM event extraction.Firstly,multi-dimensional local feature information of the text was extracted by combining convolutional neural networks with different filter window sizes,while the global feature information of the text was captured using bidirectional long short-term memory networks.Secondly,on this basis,dynamic interaction between local and global information was achieved through gating mechanisms to enhance the ability of model to identify argument boundaries.Furthermore,a fuzzy span attention mechanism was introduced to dynamically adjust the attention range,thereby optimizing the decision-making process for argument spans.Finally,label prediction was performed using conditional random fields.The results indicate that the proposed model improves the F1 score by 3.0 to 11.0 percentage points on the TCM medical records data-set,demonstrating superior performance in addressing TCM event extraction issues compared to related models.The proposed model effectively leverages both local and global semantic information of the text,enhances the flexibility of span learning and improves the capability of the model to identify argument boundaries,thereby achieving better performance in TCM event extraction.It has reference value for the inheritance and development of TCM knowledge.

关 键 词：自然语言处理 事件抽取 中医医案 注意力机制 卷积神经网络 动态融合 跨度 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] R241[自动化与计算机技术—计算机科学与技术]