In theory, that sounds a bit dangerous. If there was a good model for predicting "success" from application text, then (a) the application can be gamed, and (b) you'd probably be picking up "hot" areas that reflect the business environment at the time more than the founders. YC seems to be more about humans, which makes a strong case for using only human judgement.
a) and b) seem both overpoweringly unlikely, so I think more likely is that they'll just look at the results to see if there's anything interesting about them. (Wouldn't you like to see the results, just to see them? I would, and I have no intention of applying to YC anytime soon.)
Initially I didn't understand why one would think they would be so stupid, and then I remembered what almost every other entity that receives applications is like.
Perhaps as indicators of the business environment. For example, perhaps your model would indicate that successful startups from the last few years would include the word "social" in them. I doubt the implication goes the other way.