The Impossibility of Search Based Training

The search problem is just impossible. There are a number of papers, including one by your group, one from me, other people that show that you basically make no progress at all. If you just try to do a search based training algorithm, when all I have is a seek to seek model and answer supervision, you need some kind of additional lexical information. So we still use this model where you don't have any explicit alignment, but you can do other stuff like you had the knuckle now, more iterative search or whatever auxiliary information you can find to try to still solve this problem. In the past, I worked on large scale stuff on Freebase. I think there,