【摘要】 为解决传统的实体关系抽取模型在中文医学文本上效果不佳的问题，提出了MedRelNet网络，采用多层次语义融合策略。MedRelNet的核心是新颖的关系融合模块（RelFuse），将关系信息融合到句子表示中，实现实体和关系的充分交互。同时，引入双向长短时记忆网络（BiLSTM），全面捕捉句子特征。实验结果显示，MedRelNet相对于基线模型在CMeIE中文医学数据集、DuIE和WebNLG通用数据集上分别取得了1.0、0.7和0.8个百分点的F1值提升，这不仅表明了MedRelNet在提取医学关系三元组方面的出色表现，还突显了其较强的泛化性能。更多还原

【Abstract】 To address the issue of traditional entity-relation extraction models performing poorly on Chinese medical texts, the MedRelNet model based on Onerel is proposed. The core of MedRelNet is a novel relation fusion module(RelFuse), which integrates relation information into sentence representations to enable thorough interactions between entities and relations. Additionally, it incorporates bidirectional Long Short-Term Memory networks(BiLSTM)to comprehensively capture sentence features. Experimental results show that MedRelNet improves F1 scores by 1, 0.7, and 0.8 percentage point over baseline models on the CMeIE, DuIE, and WebNLG datasets, especially excelling in handling complex scenarios involving multiple relations and entity overlaps.更多还原