regarding this morning question about R-Prec versus P@k (question 10 of last quiz): this is in fact clear from the usage context which is described on top of page 7 (above question 8) (and which I didn't remember).
In this set-up, k is in fact a fixed internal parameter of the system and since we draw (uniformly random) ONE among the k first answers, the probability to get a correct answer at the end is indeed the number of "correct" retrieved documents in the first k (if there are 3 documents in the first k that do correspond to the right action/category, we indeed have 3/k chance to have a final correct answer). So in this framework (top page 7), average accuracy in indeed, by construction, P@k.
And, btw, we don't have R-prec in such a setup since we don't know "R" (in the sense of R-prec): what are "THE relevant documents" for a query in the framework described on top of page 7? (maybe *ALL* the ones having the right action/category, but this is not necessary since we draw at random; so another definition could be: when the probability to select the right action/category is maximal...)
Hope it clarifies the answer. If not, don't hesitate to ask more.
