基于自适应融合技术的多模态实体对齐模型  

Multi-modal entity alignment model based on adaptive fusion technology

作　　者：任楚岚 于振坤 关超 井立志 Ren Chulan;Yu Zhenkun;Guan Chao;Jing Lizhi(School of Computer Science&Technology,Shenyang University of Chemical Technology,Shenyang 110142,China;Liaoning Provincial Key Laboratory of Intelligent Technology for Chemical Process Industry,Shenyang 110142,China)

机构地区：[1]沈阳化工大学计算机科学与技术学院,沈阳110142 [2]辽宁省化工过程工业智能化技术重点实验室,沈阳110142

出　　处：《计算机应用研究》2025年第1期100-105,共6页Application Research of Computers

基　　金：辽宁省教育厅科学研究资助项目(LJKZ0449,LJKZ0434)。

摘　　要：多模态实体对齐旨在识别由结构三元组和与实体相关的图像组成的不同的多模态知识图谱之间的等价实体。现有的多模态实体对齐的研究主要集中在多模态融合策略,忽略了模态缺失和不同模态难以融合的问题,未能充分利用多模态信息。为了解决上述问题,提出了MACEA模型,该模型使用多模态变分自编码方法主动补全缺失的模态信息,动态模态融合方法整合不同模态的信息并相互补充,模态间对比学习方法对模态间进行建模,这些方法有效解决了模态缺失与模态难以融合的问题。相比于当前基线模型,MACEA的hits@1和MRR指标分别提升了5.72%和6.78%,实验结果表明,该方法可以有效地识别出对齐实体对,具有较高的准确性和实用性。Multi-modal entity alignment aims to identify equivalent entities between different multi-modal knowledge graphs composed of structured triples and images associated with entities.The existing research on multi-modal entity alignment mainly focuses on multi-modal fusion strategies,ignoring the problems of modal imbalance and difficulty in integrating different modalities,and fails to fully utilize multi-modal information.To solve these problems,this paper proposed the MACEA model,this model used the multi-modal variational autoencoder method to actively complete the missing modal information,the dynamic modal fusion method to integrate and complement the information of different modalities,and the inter-modal contrastive learning method to model the inter-modal relations.These methods effectively solve the problems of modal missing and the difficulty in modal fusion.Compared with the baseline model,MACEA improves the hits@1 and MRR indicators by 5.72%and 6.78%,respectively.The experimental results show that the proposed method can effectively identify aligned entity pairs,with high accuracy and practicality.

关 键 词：实体对齐 知识图谱 多模态 动态融合 模态缺失 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]