There are some items up from the 2006 IA Summit (Information Architecture), up at http://iasummit.org/2006/blog/, that discuss tagging.

My issue with folksonomies over a controled vocabulary is attempting to derive useful meaning when you have ambiguity -- for instance, people might classify Mail::Procmail as any one of:

mail

(ambiguous ... e-mail or postal mail?)

email

e-mail

(different spellings of the same term ... but is this sending or receiving email? neither, it's filtering, so it's still not specific)

procmail

(too specific to be a useful category in most catalogs)

smtp

(just wrong ... it filters mail that may have been sent through SMTP, but it has nothing directly to do with SMTP)

That's not to say that there isn't use for tagging -- it serves as a balance between full text searching and a full catalogging, but it can be perverted by a few (putting in bogus tags, using abnormal terminology). Folksonomies are a relatively new concept in IA, and I think it has great potential, but it is not a perfect solution -- it removes some of the problems of full text searching (number of false positives), while removing some of the problems of catalogging (expense), but doesn't really bring in the advantage of either.

I'm waiting to see how people can start data mining folksonomies to derive ontologies:

eg, PersonA .. PersonH all use Tag1, Tag2, Tag3 on the same items that PersonI .. PersonP use Tag4, Tag5; And if PersonI .. PersonP use Tag1 on an item, then Tag1 is never assigned by PersonA .. PersonH. Therefore, we can assume that Tag1 is a homonym, with different meanings to two different groups -- And between those two groups, the intersection of Tag1/Tag2/Tag3 in one group is similar to the intersection of Tag4/Tag5 by the second group.