On the Global Self-attention Mechanism for Graph Convolutional Networks