Error correction scheme augmented with statistical and lexical learning capability, for Japanese OCR

This paper proposes an improved error-correction scheme based on word-pair grammar augmented by statistical and lexical learning capability derived from manual error correction history. This learns corrections by hand to update probability score on the confusion matrix and memorizes new strings of characters gained after error-correction as to use them as a parameter of likelihood calculation for automatic error correction. The small evaluation of this algorithm has revealed, word-pair parsing based error correction improves about 2.5% of recognition rate, confusion matrix based statistical learning improves about 1.0%, and character string learning incorporated with statistical learning improves about 1.4% to output the recognition rate more than 99%, while OCR bare recognition rate is about 95%.