Leveraging orthographic information to improve machine translation of under-resourced languages