SCI-MKGC: A MKGC Method Based on Spatial Context and Interaction Attention

Multi-modal knowledge graphs (MKGs) contain diverse modalities of feature, such as text, audio and image. Due to the emergence of novel entities or an insufficiency in multi-modal corpora, MKGs confront data incompleteness frequently. Thus, MKGs completion (MKGC) is designed to predict missing entities and relations by extracting supplemental multi-model information. However, existing MKGC approaches mainly rely on embedding models and generative models, resulting in the neglect of spatial context and cross-modal interaction. To address the aforementioned challenges, we proposed a novel model named SCI-MKGC, which is a MKGC method based on spatial context enhancement and cross-modal interaction attention. In this paper, the spatial context enhancement module is proposed to enhance the central entity through aggregating neighborhood attributes. Furthermore, the cross-modal interaction attention module is designed to capture interactive semantics by cross-modal attention. The model learns discriminative semantic representations within a contrastive training framework. Experimental results demonstrate the superiority of our method on three knowledge graph inference datasets.