Analogical Reasoning Enhanced Multimodal Knowledge Graph Embedding

Multimodal knowledge graph embedding(MKGE) maps entities and relations from multiple modalities to a continuous vector space to identify potential missing triples in multimodal knowledge graphs. Most existing studies on multimodal knowledge graph embedding belong to inductive reasoning. These methods are limited in their ability to infer incomplete triples and difficult to capture complex correlations. To solve this problem, we adapt the method of analogical reasoning to implement the MKGE model enhanced by analogical reasoning. Analogical reasoning relies on the similarity between nodes for reasoning without extensive data summarization, which can compensate for the shortcomings of inductive reasoning. Specifically, this paper utilizes the vector similarity of nodes to obtain analogy nodes at three levels: relation-level, entity-level, and triple-level. Then, we design a new analogy mapping method. We generate aggregate scores based on corresponding scores using three levels of analogy nodes. Finally, we combine the original embedding method to obtain the results. We conduct experiments on WN9-IMG. The results show that our method achieves excellent performance compared to the baseline model.