What's the difference compared to just taking an embedding and feed forward a simple net trained for the task?