The power law distribution described by Zipf's law applies not only to individual words (unigrams) but also to sequences of words, such as bigrams and trigrams, though typically with a smaller exponent $\alpha$. Because many $n$-grams occur very rarely in a corpus, methods relying solely on counting statistics face a significant sparsity problem and tend to overestimate the frequency of the infrequent combinations. This high frequency of rare occurrences makes counting-based methods inadequate for language modeling and strongly motivates the transition to deep learning models.