Fitting some of the parameters of the model seems dependent on having
access to ground-truth data. An analysis of results using a combined or
out-of-sample parameter setting seems like the most appropriate, missing
test. (For example, train on 11 Wikipedia categories to pick alpha, beta,
and then try on the 12th.) This would make explicit how to select
parameters without ground-truth, or with only partial, ground-truth.