Towards Zero-shot Learning for End-to-end Cross-modal Translation Models F i n d i n g s - E M N L P