At that point, you do need some sort of grounding, e.g. a human reading the text, or a neural network embedded in a robot. It isn't symbols all the way down.