Characterizing the knowns and unknowns of text simplification models