After doing translation of some GIMP docs in gimp.ru translation team i have
come to conclusion that the current framework is not efficient in comparison to
e.g. KDE translation framework.

Reasons are:

1. All translations of a documentation item are stored in one file.

This is inefficient because it means that there are at least 6 teams working on
translation of the very same file, so it's difficult to stay tuned to all
changes they introduce.

Versioning becomes specific language independent too and thus a) gives no idea
of progress for a specific language, b) makes it pretty much difficult to
compare different versions of a translation into only one language (xml diffs
are not human readable really).

2. There is no way to automate merging changes into translations. 

KDE translation team efectively uses xml2pot service, which means:

a) the author of an English documentation finishes another iteration of his work;
b) he runs a script that translates his xml file to a pot messages catalog file
and updates all the rest *.po, so that new sentences appear as untranslated and
changes sentences -- as fuzzy;
c) the author commits all these changes;
d) translators update from CVS, then update translation of these *.po with
KBabel, GTranslator, poEdit or whatever they use and commit these changes back
to CVS;
e) when it comes to the next release of an application, another script is run
over *.po to create new xml files from them.

Thus 

a) instead of having awful headache with xml diffs and text editors translators
use neat tools they got used to;
b) merging changes becomes automated.

I have 4 years of experience in translation of open source applications -- just
in case you think that I don't know what I'm talking about ;)