摘要
为解决实体关系抽取任务中的关系重叠问题,提出一种融合图注意力网络(graph attention networks, GAT)的层级标注联合抽取方法。将关系和词建模为图结构上的节点,通过GAT的“消息传递”机制实现两类语义节点信息传递、融合与更新,实现两类节点间的完整信息交互,在标注阶段,采用层级标注策略,解决关系重叠问题,使用Focal Loss损失函数对模型进行训练,缓解标注阶段数据不均衡的问题。实验结果表明,该方法具有良好的性能,能够高效抽取出重叠关系三元组。
To solve the problem of overlapping relations in joint entity and relation extraction,a hierarchical annotation joint extraction method was proposed.Entities and relationships were modeled as nodes on a graph structure,and the message passing mechanism of graph attention networks(GAT)was used to achieve the complete information interaction between entities and relationships by transmitting and integrating two types of semantic node information.In the tagging stage,a hierarchical tagging strategy was applied to effectively solve the problem of relationship overlap,and the Focal Loss function was used to train the model,alleviating the problem of data imbalance in the tagging stage.Experimental results show that the proposed method has good performances and can extract overlapping relationship triplets efficiently.
作者
蔡阿雨
黄洁
张克
CAI A-yu;HUANG Jie;ZHANG Ke(School of Cyber Science and Engineering,Zhengzhou University,Zhengzhou 450003,China;College of Data and Object Engineering,Information Engineering University,Zhengzhou 450001,China)
出处
《计算机工程与设计》
北大核心
2025年第5期1378-1386,共9页
Computer Engineering and Design
基金
国家自然科学基金项目(62071490)
河南省自然科学优秀青年基金项目(212300410095)。
关键词
联合抽取
关系重叠
图结构
图注意力网络
层级标注
消息传递
损失函数
joint extraction
overlapping relations
graph structure
graph attention networks
hierarchical annotation
message passing
loss function
