Subtract a small fixed amount from all observed (non-zero) counts $d$. In effect, this reserves some mass for the unseen items. Empirically estimate $d$ by splitting the corpus in two equal parts and observing how often on average do bigrams occur in only one of the two halves.