One can find lots of reasons why it can't possibly work. All those reasons would also lead one to believe that a child could not possibly learn to speak or to understand. And yet children do learn from all of this mess. And it just seems to be the case that neural networks do the same thing. Given enough text, they might be able to infer everything that's missing. It's a dirty approach. But it's the only one that has produced the kind of results that lets people say, "Oh wow!".