Effectively Filtering Images for Better Multi-modal Knowledge Graph

Huang Peng, Hao Xu, Jiuyang Tang, Jibing Wu, Hongbin Huang

Abstract: The existing multi-modal knowledge graph construction techniques have become mature for processing text modal data, but lack effective processing methods for other modal data such as visual modal data. Therefore, the focus of multi-modal knowledge graph construction lies in image and image and text fusion processing. At present, the construction of multi-modal knowledge graph often does not filter the image quality, and there are noises and similar repetitive images in the image set. To solve this problem, this paper studies the quality control and screening of images in the construction process of multi-modal knowledge graph, and proposes an image refining framework of multi-modal knowledge graph, which is divided into three modules. The final experiment proves that this framework can provide higher quality images for multi-modal knowledge graphs, and in the benchmark task of multi-modal entity alignment, the effect of entity alignment based on the multi-modal knowledge graphs constructed in this paper has been improved compared with previous models.