Cross Attention Scoring Function for Multimedia Domain Knowledge Graph Completion

Multimedia-based applications and services are strongly expanded by the continuous development of communication and network technologies. Recently, with the high efficiency of knowledge graph(KG) in data modeling, various KGs in multimedia domains have been constructed to manage and utilize multimedia services. In multimedia KGs, the integrity of the link structure can significantly impact the performance of downstream applications. Numerous knowledge graph completion(KGC) models are proposed to predict lost relations between KG entities. However, these general models target at capturing different relation patterns and lack the attention of the relationship between entities and relations. In this paper, a novel KGC model named TransX is proposed to learn this complex inner relationship, which can robustly reflect the link structure of the multimedia domain KGs. TransX firstly projects entities and relations to multiple property spaces, and then a new cross attention block is designed to represent different patterns between entity and relation properties. The experiment results demonstrate that TransX has fewer parameters and better performance at the cost of introducing some sufferable computation in the cross attention block.