MULTIFORM: Few-Shot Knowledge Graph Completion via Multi-modal Contexts

Abstract

Knowledge Graphs (KGs) have been applied to many downstream applications such as semantic web, recommender systems, and natural language processing. Previous research on Knowledge Graph Completion (KGC) usually requires a large number of training instances for each relation. However, considering the accelerated growth of online information, there can be some relations that do not have enough training examples. In fact, in most real-world knowledge graph datasets, instance frequency obeys a long-tail distribution. Existing knowledge embedding approaches suffer from the lack of training instances. One approach to alleviating this issue is to incorporate few-shot learning. Despite the progress they bring, they sorely depend on entities’ local graph structure and ignore the multi-modal contexts, which could make up for the lack of training information in the few-shot scenario. To this end, we propose a multi-modal few-shot relational learning framework, which utilizes the entities’ multi-modal contexts to connect few instances to the knowledge graphs. For the first stage, we encode entities’ images, text descriptions, and neighborhoods to acquire well-learned entity representations. In the second stage, our framework learns a matching metric to match the query triples with few-shot reference examples. The experimental results on two newly constructed datasets show the superiority of our framework against various baselines.