Statistical methods can deal with black swans, but you've got to get
away from normal distributions and also model the risk that your model is
wrong.
Since training sets come from the same place sausage comes from,
training sets in machine learning rarely teach the algorithm the correct
prior distribution of the class. Punch a new prior into the system and it
will perform much better.
Some kinds of sampling biases can be somewhat overcome. Involvement of
multiple people smoothes out individual bias. (Kurzweil's project of
stealing a human soul with a neural network is already being scoops by
projects that are stealing statistical models of many souls.)
Language zone Wikipedias are obviously biased towards the viewpoint of
people in that language zone. Mostly that's a good thing, because a
Chinese knowledge base that reflected an Anglophone bias would seem
unnatural to Chinese speakers.
And that's the point. Useful systems don't "eliminate bias" but
they
are given the bias that they need in order to do their job.
I agree categories are most useful when they are the categories you
need. The toolbox above can help you estimate these with precision so high
that it's difficult to measure.
Arnold S isn't the best case for categories because humans,
bodybuilders, places, chemicals and such are well ontologized. Look at
the collection that comes up for the word "Intersection",
http://en.wikipedia.org/wiki/Intersection
Most of these are connected to the larger mass through just a few
categories that would be hard to express as restriction types. Wikipedia
is reasonable to require concepts to have a category because really, if you
want to assert something exists and can't find some category that this thing
is a member of, I wouldn't be so sure that this thing exists.
I'm not sure if there is anything I can't do with the current situation,
but bear in mind that I'm going to look at DBpedia, Wikidata and Freebase
facts too and be willing to do data cleaning processing and hand cleaning of
results that I cannot accept. It's a tricky and somewhat expensive process
(though it's cheaper than conventional ontology construction), so cleaner
data makes this process cheaper and quicker and available to more end users
personalized to their own needs to define the categories they need.