Tackling Ambiguity with Images: Improved Multimodal Machine Translation and Contrastive Evaluation