Zipf's law observes that the distribution of frequencies of linguistic sequences has a long tail. Formally $Count(w_i) \propto \frac{1}{Rank(w_i)}$ This simply means that data are sparse for most occurrences.