摘要
【目的】研究适用于科技文献的机构名称对齐方法。【方法】利用大语言模型并结合提示工程,从科技文献著录信息中识别机构及其关联实体,构建机构知识图谱,通过文本嵌入和图卷积网络,基于所构建的知识图谱实现同一机构不同表述形式的名称对齐。【结果】与直接采用文本对齐的方法相比,论文提出的方法在无需要额外训练的情况下,机构名称对齐性能指标Hit@1、Hit@10和MRR分别提升约24%、7%和11%。【局限】处理复杂信息结构文本中的跨语种、多模态的实体对齐效果有待提升。【结论】融合提示工程与图卷积网络的机构名称对齐方法,可有效提升科研机构名称规范库建设的质量与效率。
[Objective]This paper proposes an institution name alignment method for scientific literature.[Methods]By combining large language models with prompt engineering,we identified institution names and their related entities from bibliographic information to construct an institutional knowledge graph.Then,we utilized text embedding and graph convolutional networks(GCNs)to align different expression forms of the same institutional name based on the constructed knowledge graph.[Results]Compared with direct text-based alignment methods,our approach improved alignment performance without requiring additional training.Specifically,Hit@1,Hit@10,and MRR scores are improved by approximately 24%,7%,and 11%,respectively.[Limitations]The method still faces challenges in aligning cross-lingual and multimodal entities in texts with complex information structures.[Conclusions]The proposed method can enhance the efficiency and effectiveness of building standardized scientific research entity name repositories.
作者
王茜
方安
娄培
杨雨生
王蕾
Wang Qian;Fang An;Lou Pei;Yang Yusheng;Wang Lei(Institute of Medical Information,Chinese Academy of Medical Sciences,Beijing 100020,China)
出处
《数据分析与知识发现》
北大核心
2025年第8期125-136,共12页
Data Analysis and Knowledge Discovery
基金
2023年国家科技图书文献中心专项课题医学领域规范数据加工和丰富化(项目编号:2023XM2401)的研究成果之一。
关键词
实体识别
实体对齐
提示工程
知识图谱
图卷积网络
Entity Recognition
Entity Alignment
Prompt Engineering
Knowledge Graph
Graph Convolution Networks(GCN)
