I left Google in 2010, so it's just a wild guess, but I suspect a big part of the issue is learn-to-rank is probably being trained on everyone's searches. I think it would probably do much better if they used the presence or absence of search operators as a simple heuristic to separate power user searches from common searches, and trained a separate ranking model on power user searches.
Maybe they're already doing this, but it sure acts like learn-to-rank is always ranking pages as if the query were very sloppy.
It's been a long time, and I certainly never read the code, but I vaguely remember a Google colleague mentioning something (before learn-to-rank) about a back-end query optimizer branch that would intentionally disable much of the query munging if there were any search operators in the query. There was some mention about using cookies / user account information to do the same if the same browser/user had used any search operators in the past N days, but I'm not sure if that was implemented or just being floated as a useful optimization.
Maybe they're already doing this, but it sure acts like learn-to-rank is always ranking pages as if the query were very sloppy.
It's been a long time, and I certainly never read the code, but I vaguely remember a Google colleague mentioning something (before learn-to-rank) about a back-end query optimizer branch that would intentionally disable much of the query munging if there were any search operators in the query. There was some mention about using cookies / user account information to do the same if the same browser/user had used any search operators in the past N days, but I'm not sure if that was implemented or just being floated as a useful optimization.