ngrambased

N-gram based methods are statistical approaches in natural language processing that use contiguous sequences of n items—words or characters—to model language. An n-gram model estimates the probability of a text by decomposing it into a chain of conditional probabilities, where each token depends on the previous n−1 tokens. Training relies on a corpus to count occurrences and derive frequency-based estimates.

Word-level and character-level variants are common. Word n-grams capture lexical sequences, while character n-grams capture subword patterns and handle unseen words. Probabilities P(wi | w_{i-n+1} ... w_{i-1}) are estimated from counts, often with smoothing (Laplace, Good-Turing, or backoff/interpolation) to handle unseen n-grams. Larger n encodes more context but needs more data and memory; smaller n is simpler but less expressive.

Applications include predictive typing, spelling correction, language identification, information retrieval, and as features in classifiers. They served as a foundational baseline for statistical NLP and remain useful as lightweight components or in hybrid systems. Limitations include inability to model long-range dependencies beyond n and data sparsity for large n, as well as shallow semantics. In modern NLP, neural language models provide alternatives, though n-gram methods persist in resource-constrained or specialized settings.