Improving language modeling for (off-line) Chinese character recognition

Abstract

We analyze the error characteristics of a Chinese character recognizer and developed two approaches to improve Chinese character recognition system. We first develop a non-contiguous context dependent language model as a post processing module. The model makes use of far away context to predict the interested character. The model is only as good as the traditional bigram model in terms of accuracy. Secondly, we developed a method to detect errors in language model. The method employs pattern recognition technique. It combines both dictionary and statistical features to predict whether a block of character is correct or contains error. This detection scheme as demonstrated in our experiment is effective. The performance is 80%, 91% and 75% of precision, recall and skip ratio respectively.