At least searching and ranking a mere 75,000 documents to find good employees shouldn't be a problem for them.
To give just the easiest and most recent example of how that approach breaks: the increasing deluge of content farms and SEO spam in Google results. Many folks consider that non-optimal but Google defends it because their data "tells them" it's optimal. Now picture that approach with people. Rank things by some measurement and only things that measure well by that metric will rank high, shutting you out of great opportunities. The false negative, etc. Related anti-pattern is the concept of Local Maxima.
C'mon, I'm not one to usually come rushing to Google's aid, but from Matt Cutts' comments on HN and recent blog posts on the official blog is seems to me that they are working hard on this problem, which would imply they don't think the status quo is optimal.
Secondly, it should be possible to do something as simple as filter out obvious content farm sites in under a week tops. Or give all users a quick and easy to do it. Perhaps have a default blacklist of known content farmers, and then give users the option to remove things from the blacklist when doing their own searches. You might be dealing with a huge company if something as simple as this takes 20k employees more than a week, tops, to put into production. And I'm being generous with the week estimate, because it could be done even faster than that under more agile conditions. And a 'default blacklist of known content farmers' is not even the only solution which could be implemented quickly, you could also do something like add another variable to the PageRank algorithm which has a down weighing effect on any page under a domain known to be a content farmer/spammer. Supposedly, they are already doing this (IIRC) but again, the basic idea could be implemented and put into production quickly, if it hasn't already.
At the very least, it would allow you to filter the chaff, and enough care could be given to a dataset as small as 75,000 documents to take care of preprocessing, some manual curation, etc. and yield some decent results. If you wanted to, you could even compute PageRank scores in R on your laptop for a dataset that small (and that would probably be the only computationally intensive part of it, after POS tagging and ML model fitting).
As a bonus, the effort would go a long way to help your recruiting in the future.