多级融合知识图谱补全模型  

Multi-level Fusion Knowledge Graph Completion Model

作　　者：叶志鸿 吴运兵[1] 戴思翀 曾智宏 YE Zhihong;WU Yunbing;DAI Sichong;ZENG Zhihong(College of Computer and Big Data,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学计算机与大数据学院,福州350108

出　　处：《计算机科学与探索》2025年第3期724-737,共14页Journal of Frontiers of Computer Science and Technology

基　　金：福建省自然科学基金(2022J01116);福建省高等教育改革与研究项目(FGJG202411)。

摘　　要：知识图谱补全旨在通过预测缺失的三元组来扩展和完善知识图谱,多模态知识图谱补全融合了实体的本体信息,如实体描述、实体图像和实体属性,以获取更精确的实体表示。现有研究将不同模态投影到统一的空间中,以获取实体模态联合表示,再融合知识图谱结构信息作出预测。然而,现存方法融合多模态信息时难以捕捉实体背景知识的复杂交互,不可避免地存在信息丢失和特征提取能力不足的问题;同时过拟合及实体关系交互不足限制了二维卷积模型性能,导致难以融合知识图谱结构信息。因此,提出了多级融合知识图谱补全模型,从实体多模态信息融合与知识图谱结构信息融合两方面解决上述问题。为充分融合实体多模态信息,提出同时使用三种不同融合方法,以全面捕捉实体背景知识交互,并联合决策学习,旨在结合不同多模态融合方法提供的互补信息,以获取实体丰富多样的表示;为充分融合知识图谱结构信息,利用特征泛化来缓解二维卷积模型的过拟合问题,并结合特征重塑增强实体与关系间交互,以提升实体与关系间的上下文感知能力。实验结果表明,该模型在多个公开数据集上均取得较好性能。Knowledge graph completion aims to expand and enhance knowledge graphs by predicting missing triples.Multi-modal knowledge graph completion integrates entity ontology information such as entity descriptions,entity images,and entity attributes to obtain more accurate entity representations.Existing research projects different modalities into a unified space to obtain joint representations of entities,then combine knowledge graph structural information for predictions.However,existing methods have difficulty in capturing the complex interactions between entity background knowledge when fusing multi-modal information,which inevitably leads to information loss and insufficient feature extraction capabilities;overfitting and limited entity relation interactions restrict the performance of 2D convolution models,making it difficult to integrate knowledge graph structural information.Therefore,this paper uses a multi-level fusion knowledge graph completion model to address the above issues from two aspects:the fusion of entity multimodal information and the integration of knowledge graph structural information.To fully integrate entity multi-modal information,three different fusion methods are simultaneously used to comprehensively capture the interaction of entity background knowledge,along with decision learning,aiming to combine the complementary information provided by different multi-modal fusion methods to obtain rich and diverse entity representations.To fully integrate knowledge graph structural information,feature generalization is proposed to alleviate the overfitting issues of 2D convolution models,combined with feature reshaping to enhance interactions between entities and relations,thereby improving the contextual perception ability of entities and relations.Experiments on multiple public datasets demonstrate the superior performance of the proposed method.

关 键 词：知识图谱补全 多模态融合 本体信息 结构信息 决策学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]