Revisiting Document Representations for Large-Scale Zero-Shot Learning