融合图像信息的多嵌入表示实体对齐方法

Multi-Embedding Representation Entity Alignment Method Based on Image Fusion Information

实体对齐是知识图谱融合技术的关键环节,然而现有方法在处理跨语言图谱时未能充分利用图谱数据,在此提出一种方法融合图像信息的多嵌入表示实体对齐方法。该方法从不同角度获取文本嵌入,并利用图像数据丰富文本嵌入,实现多模态信息融合以完成跨语言图谱的实体对齐任务。通过图像生成模型解决实体图像覆盖不完全问题,结合迭代策略获得高质量实体图像信息以扩充跨语言知识图谱中种子序列对。为了更好适用现实世界真实知识图谱融合过程,该方法将对齐阶段转换为二分图匹配问题。提出的方法在公开数据集上进行了实验分析,实验结果表明了方法的良好性能,还通过消融实验验证各模块的有效性,并针对不同情况提供了参数的可选择性。

Entity alignment is the key step of knowledge graph fusion technology.However,existing methods fail to make full use of graph data when processing cross-language graph.Hence,the paper proposes a multi-embedding repre-sentation entity alignment method based on image fusion.This method obtains text embeddings from different angles,enriches text embeddings with image data,and realizes multi-modal information fusion to complete entity alignment across linguistic knowledge graph.The image generation model is used to solve the problem of incomplete entity image coverage,and the high-quality entity image information is obtained by the iterative strategy to expand the seed sequence pairs in the cross-language knowledge graph.In order to better apply the knowledge graph fusion process in the real world,the method transforms the alignment phase into a binary graph matching problem.The proposed method is experi-mentally analyzed on a public data set,and the experimental results show the good performance of the method.The ablation experiment also verifies the effectiveness of each module,and provides the parameter selectivity for different situations.