Attention-Based Aggregation Graph Networks for Knowledge Graph Information Transfer

Knowledge graph completion (KGC) aims to predict missing information in a knowledge graph. Many existing embedding-based KGC models solve the Out-of-knowledge-graph (OOKG) entity problem (also known as zero-shot entity problem) by utilizing textual information resources such as descriptions and types. However, few works utilize the extra structural information to generate embeddings. In this paper, we propose a new zero-shot scenario: how to acquire the embedding vector of a relation that is not observed at training time. Our work uses a convolutional transition and attention-based aggregation graph neural network to solve both the OOKG entity problem and the new OOKG relation problem without retraining, regarding the structural neighbors as the auxiliary information. The experimental results show the effectiveness of our proposed models in solving the OOKG relation problem. For the OOKG entity problem, our model performs better than the previous GNN-based model by 23.9% in NELL-995-Tail dataset.