Unigrammalli

Unigrammalli, or unigram model, is a simple probabilistic language model that treats each word as generated independently from a fixed probability distribution over the vocabulary. In practice, the probability of a word sequence w1 w2 ... wn is approximated as the product P(w1) P(w2) ... P(wn). The model ignores word order and context, hence it is a baseline rather than a full language model.

The probabilities P(w) are typically estimated from a corpus as relative frequencies P(w) = count(w) / N, where N is the total number of word tokens. This is the maximum likelihood estimate (MLE). To handle unseen words or to improve robustness, smoothing techniques such as Laplace (add-one), Good-Turing, or Kneser-Ney are used.

Unigram models serve as simple baselines in various NLP tasks. They are used in text classification, language identification, spelling correction, and as a component in information retrieval under the query likelihood framework. They can also be used as a starting point for more complex n-gram models or as a feature in machine learning pipelines.

Limitations include the strong independence assumption, which ignores syntax, semantics, and word order. Consequently, unigram models often have higher perplexity on natural language tasks compared to higher-order n-gram models. Despite limitations, they are computationally lightweight and provide interpretable estimates, making them useful for quick assessments and baseline comparisons.