Adaptive construction of a statistical language model

A method of creating a language model, the method comprising:

• using a processor to perform acts comprising:
• analyzing text in a set of documents to determine a set of probabilities associated with N-grams observed in said set of documents;
• calculating a first quantity that represents how well a first statistical language model predicts occurrence of said N-grams in said set of documents;
• calculating a second quantity based on said first quantity;
• creating a second statistical language model based on said second quantity, said first statistical language model, and said set of probabilities,
• said second quantity comprising a probability that said first statistical language model correctly predicts occurrence of said N-grams in said set of documents, and said creating of said second statistical language model comprising:
• applying said second quantity to said first statistical language model;
• and
• applying a complement of said second quantity to said set of probabilities.

A system for creating a language model, the system comprising:

• a processor;
• a data remembrance component;
• a document crawler that evaluates documents and that counts occurrence of N-grams in said documents and that creates data indicating counts of N-grams in said documents;
• and
• a language model builder that receives data, from said document crawler, that indicates counts of N-grams in a first set of documents, wherein said language model builder creates a quantity indicating how well a first statistical language model predicts counts of N-grams in said first set of documents, and creates a second statistical language model based on a first probability distribution in said first statistical language model, a second probability distribution based on N-gram counts of said first set of documents, and said quantity, said quantity comprising a probability that said first statistical language model correctly predicts occurrence of said N-grams in said set of documents, said language model builder creating said second statistical language model by applying said quantity to said first statistical language model and by applying a complement of said quantity to a set of probabilities derived from N-gram counts in said first set of documents.