GMatchBERT: BERT-Enhanced Relation Learning for Knowledge Graphs

Learning the relations between two items is an important step in completing the knowledge graph. Knowledge graph completion typically requires numerous training examples for each relation to increase their coverage. In KGs, long-tail relations are more prevalent, and freshly inserted links generally don’t have enough training triplets. One-shot relation learning is a transformational approach in the field of knowledge graphs, tackling the issues given by data scarcity and the necessity for fast reasoning mechanisms. Existing one-shot relation learning approaches rely on embedding-based methods based on the continuous vector-space. This study proposes GMatchBERT, a one-shot relation learning method with BERT embedding. The approach uses information from BERT embeddings and integrates the embeddings with one-hop graph topologies to create a matching metric. Experiments on two sets of data show that our suggested framework surpasses the existing approach for one-shot relation learning.