I assume we're talking about Cohen's kappa here.
https://en.wikipedia.org/wiki/Cohen's_kappa

--------------------------------------------------
Note added at 1 hr (2015-08-27 18:04:39 GMT)
--------------------------------------------------

The tau here may be that of Goodman and Kruskal. It gives the probabilities of assigning cases correctly to a set of categories, in the light of data on another set of categories.

--------------------------------------------------
Note added at 5 hrs (2015-08-27 21:54:46 GMT)
--------------------------------------------------

The tau coefficient used here is, I think, that due to Ma and Redmond (and NOT Kruskall or Kendall!) for a definition see http://www.unesco.org/csi/pub/source/rs9.htm

This is essentially the same as Naesset's "Tau baseia-se na probabilidade a priori (NAESSET, 1996), ou seja, a concordância esperada (Pr) pode ser obtida antes mesmo de elaborar a matriz de erros. P r = 1/k, onde k é número de categorias ou classes."

So this would be called the "tau index of agreement" and the kappa is the "index of interrater reliability".