Learning visually grounded words and syntax for a scene description task