From route descriptions to sketches: a model for a text-to-image translator