Hi Shaswata, N should be the total # of unique words in the corpus. The derivation is for illustration purpose only assuming we are calculating the perplexity of the entire corpus. Usually, we don’t assume all words have the same probability 1/N.