Magnus:
I've asked this before, and got some vaguely positive answers, so I'll try again...
Ah yes, that sounds familiar: http://mail.wikimedia.org/pipermail/wikitech-l/2004-March/021553.html ;-)
As Gerard has pointed out, I'm working on Wikidata. My current line of thinking is at http://meta.wikimedia.org/wiki/Wikidata/Notes (not really meant for public consumption yet, but feel free to point out obvious stupidities).
Wikidata does not necessarily require a central repository, you can have project-local or global schemas, though in most cases, having data centrally stored will be preferable. Take, for example, country statistics: The population, the currency, the GDP, the head of state etc. are not language-dependent. You will not want to have redundant copies of all this data or you end up with a high maintenance system like the interlanguage links. ;-) Instead you will want to be able to distinguish between * data which is highly language-specific (leave it blank even if data in other languages exists) * data which is rarely language-specific (make a copy from other language(s)) * data which is never language-specific (everyone edits the same version). Ideally, you would be able to group subsets of languages together according to these criteria, and define orders of preference.
The user does not necessarily have to notice that they are using a global repository. You could have the script access the DB transparently as needed. But all this is still quite a while away, the Ultimate Wiktionary is priority number 1, and that's a separate database of its own.
When I've finished the specs (I may have to go back to the drawing board if the current fixed table approach turns out to be unworkable), I'll send out a call for open participation. I'd like to ask you to be patient until then with your own implementation. As for importing existing data, we could use a server-side script to extract data from templates, though this should be done cautiously as much of it is currently not very high quality (templates have no parameter validation).
Erik