Challenging the Assumption of Structure-based embeddings in Few- and Zero-shot Knowledge Graph Completion