Query spelling correction is a crucial component of moden search engines that can help users to express an information need more accurately and thus improve search quality. In participation of the Microsoft Speller Challenge, we proposed and implemented an efficient end-to-end speller correction system, namely CloudSpeller. The CloudSpeller system uses a Hidden Markov model to effectively model major types of spelling errors in a unified framework, in which we integrate a large-scale lexicon constructed using Wikipedia, an error model trained from high confidence correction pairs, and the Microsoft Web N-gram service. Our system achieves excellent performance on two search query spelling correction datasets, reaching 0.970 and 0.940 F1 scores on the TREC dataset and the MSN dataset respectively.

This paper presents an integrated, end-to-end approach to online spelling correction for text input. Online spelling correction refers to the spelling correction as you type, as opposed to post-editing. The online scenario is particularly important for languages that routinely use transliteration-based text input methods, such as Chinese and Japanese, because the desired target characters cannot be input at all unless they are in the list of candidates provided by an input method, and spelling errors prevent them from appearing in the list. For example, a user might type suesheng by mistake to mean xuesheng 学 生 'student ' in Chinese; existing input methods fail to convert this misspelled input to the desired target Chinese characters. In this paper, we propose a unified approach to the problem of spelling correction and transliteration-based character conversion using an approach inspired by the phrasebased statistical machine translation framework. At the phrase (substring) level, k most probable pinyin (Romanized Chinese) corrections are generated using a monotone decoder; at the sentence level, input pinyin strings are directly transliterated into target Chinese characters by a decoder using a loglinear model that refer to the features of both levels. A new method of automatically deriving parallel training data from user keystroke logs is also presented. Experiments on Chinese pinyin conversion show that our integrated method reduces the character error rate by 20 % (from 8.9 % to 7.12%) over the previous state-of-the art based on a noisy channel model. 1