qSpell: Spelling Correction of Web Search Queries using Ranking Models and Iterative Correction

In this work we address the challenging problem of Web search queries in order to build a speller that proposes the most plausible spelling alternatives for each query. First we generate a large set of candidates using diverse approaches including enumerating all possible candidates in edit distance of one, fuzzy search on known data sets, and word breaking. Then, we extract about 150 features for each query-candidate pair and train a ranking model to order the candidates such that the best candidates are ranked on the top of the list. We show that re-ranking top results, iterative correction, and post-processing of the results can significantly increase the precision of the spell checker. The final spell checker, named qSpell, achieves a precision of 0.9482 on a test data set randomly collected from real search queries. qSpell won the 3rd prize in the recent Microsoft’s Speller Challenge.