Knowledge Graph Completion Algorithm Based On Multimodal Representation Learning

In the existing multi-modal knowledge graph, the entity picture datasets have irrelevant pictures, resulting in noise of knowledge representation learning. Therefore, this paper proposes to use the image filtering algorithm to remove the entity irrelevant pictures in the picture datasets to improve the quality of datasets, and use the image enhancement algorithm to improve the number of pictures in the picture datasets. In this paper, the pretrained-model Bert was used to extract the text features, the pretrained-model CLIP was used to extract the image features, and the modal fusion algorithm was used to get the final representation of the triples. In addition, for textual and visual feature fusion, this paper proposes a multi-lavel residual network features fusion algorithm, and uses the public dataset FB15K-237 to test the proposed knowledge completion algorithm (MMKGR) based on multimodal knowledge representation learning. After comparing with other multimodal knowledge completion algorithms,as TransE, VisualBERT_base,IKRL and TransAE the algorithm proposed in this paper is superior to other algorithms in Hit@1, Hit@10 and MR.