DM-MKGC: Multimodal Knowledge Graph Completion Based on Dynamic Prompt Learning and Multi-Granularity Aggregation

Multimodal knowledge graph completion (MKGC) seeks to enrich knowledge graphs by integrating information from diverse modalities, facilitating more comprehensive knowledge representation and enhancing reasoning accuracy. However, existing models lack the flexibility to adapt to different tasks, and their performance still requires further improvement. To tackle these challenges, we propose an MKGC model based on Dynamic prompt learning and Multi-granularity cross-modal Aggregation, namely DM-MKGC. To be specific, a novel dynamic prompt template is proposed, which employs an adaptive task-guided mechanism to dynamically adjust the structure of entities, relations, and textual information. This approach enables the generation of prompts tailored to diverse tasks, ensuring both functional flexibility and structural consistency. Furthermore, a multi-granularity cross-modal aggregation method, which facilitates the aggregation of cross-modal information by facilitating the interaction between coarse-grained and fine-grained image features with textual features, is designed to enhance the model’s performance. Extensive experiments conducted on four datasets (FB15k-237-IMG, WN18-IMG, MNRE, and Twitter-2017) demonstrate our model outperforms other SOTA methods in knowledge completion, achieving an average improvement of 9.8%, 1.6%, and 1.23% in MR, Hits@n, and F1 respectively. Our model not only offers a novel method for multimodal knowledge graph completion but also contributes valuable insights for the advancement of knowledge graph technologies.