A bigram is a pair of consecutive words in a text sequence. In natural language processing (NLP), bigrams are used to build statistical models that estimate how likely a word is to appear based on the word immediately before it.
The model works on the assumption that you don't need to read an entire sentence to make a reasonable guess about what word comes next. Just knowing the previous word is usually enough. So rather than asking “Given everything before this point, what comes next?” a bigram model asks, “Given just this one word, what typically follows?”
Bigrams sit within the broader family of n-gram models, which extend the same idea to longer sequences—trigrams look back two words, and so on. Transformer-based models like BERT and GPT have largely replaced n-gram approaches for most NLP tasks today, but bigrams remain a useful baseline and a foundational concept for understanding how statistical language modeling works.