one observation i want to make concerns the appropriate reliability
coefficient.  you seem to have settled on kappa.  i am suggesting to you
that this is an inappropriate statistics, at least in the domain of content
analysis or the coding of qualitative data more generally, as it excludes
unreliabilities due to unequal coder preferences for categories (manifest in
unequal marginal frequencies of a contingency table).  i have written about
kappa's inconsistent definition.  in its denominator kappa conforms to
association/correlation statistics, e,g. chi square, while in its numerator
it conforms to agreement statistics, e.g. scott's pi.  this makes kappa
hybrid whose values are difficult to interprete.  

alpha includes this source of unreliability.  you may notice that when
margins are unequal, kappa is always larger than scott's pi, the difference
being due to unreliabilities hat kappa ignores but pi includes.  alpha is
not a derivative or form of kappa, as you state.  if you want to relate it
to another coefficient: when sample sizes are very large, the nominal and
two coder form of alpha equals scott's pi.  historically, pi preceeded kappa
(and alpha).  if you have large numbers of two coder and nominal data, pi is
preferrable to kappa.  in all other situations i have to recommend alpha.