Revisiting Simple Neural Networks for Learning Representations of Knowledge Graphs