Fine tuning on support vector regression parameters for personalized english word-error correction algorithm

Abstract

A better understanding on word classification and regression could lead to a better detection and correction technique. We used different features or attributes to represent a machine-printed English word, and support vector machines is used to evaluate those features into two class types of word: correct and wrong word. Our proposed support vectors model classified the words by using fewer words during the training process because those training words are to be considered as personalized words. Those wrong words could be replaced by correct words predicted by the regression process. Our results are very encouraging when compared with Microsoft's spell checker, and further improvement is in sight.