Predictive stemming for web search with statistical machine translation models

Techniques for determining when and how to transform words in a query to return the most relevant search results while minimizing computational overhead are provided. A dictionary is generated based upon words used in a specified number of previous most frequent search queries and comprises lists of transformations that may include variants based upon the stems of words, synonyms, and abbreviation expansions. When a query is received from a user, candidate queries are generated based upon replacing particular words in the query with a transformation of the particular words. Candidate queries are selected that have a high probability of returning relevant results by computing values of the query using language model scoring and translation scoring. The selected candidate queries and the original query are executed to return search results. The search results are displayed to the user with the words in the original query and the transformed words in bold.