The Application of Machine Transliteration Techniques to Spelling Correction
Keiko Taguchi, Andrew Finch, Seiichi Yamamoto, Eiichiro Sumita · 2015
This paper extends existing work on spelling correction using statistical machine translation by incorporating techniques that have proved valuable in the related field of machine transliteration. We investigate training the models using a non-parametric Bayesian aligner, alternative translation model features, a language model trained to bias the decoding process towards producing words from a dictionary, and the integration of a joint source-channel model into the set of log-linear models. Our experiments show that all of the enhancements we propose can match or improve the accuracy over a respectable baseline phrase-based statistical machine translation system. Furthermore, the Bayesian aligner gave rise to considerably more compact models and the proposed language model results in a more efficient decoding process by eliminating partial hypotheses that cannot lead to useful results from the search graph.