Learning Zero-Shot Multifaceted Visually Grounded Word Embeddings via Multi-Task Training