Candidate words in search queries are identified, each candidate word including one or more consecutive characters. For each candidate word, a first count is determined, the first count representing a number of times that the candidate word is the only word in the search queries, and a second count is determined, the second count representing a number of times that the candidate word and one or more other words are included in each of the search queries. One or more of the candidate words are added to an input method editor dictionary based on a relationship between the first count and the second count.