Comment by elandau25

That's a good idea. It will likely be trickier to apply a method like this for text. Decomposition of the problem is less obvious than with vision tasks.