MMKRL: A robust embedding approach for multi-modal knowledge graph representation learning

摘要

Most knowledge representation learning (KRL) methods only use structured knowledge graphs (KGs); however, there is still much multi-modal (textual, visual) knowledge that has not been used. To address this challenge, we propose a novel solution called multi-modal knowledge representation learning (MMKRL) to take advantage of multi-source (structured, textual, and visual) knowledge. Instead of simply integrating multi-modal knowledge with structured knowledge in a unified space, we introduce a component alignment scheme and combine it with translation methods to accomplish multi-modal KRL. Specifically, MMKRL firstly reconstructs multi-source knowledge by summing different plausibility functions and then aligns multi-source knowledge using specific norm constraints to reduce reconstruction errors. We also select an adversarial training strategy to enhance the robustness of MMKRL, which is rarely considered in existing multi-modal KRL methods. Experimental results show that MMKRL can effectively utilize multi-modal knowledge to achieve better link prediction and triple classification than other baselines in two widely used datasets. Further, when relying on structured knowledge or limited multi-source knowledge, MMKRL still achieves competitive results in link prediction, demonstrating our model’s superiority.