A Contextual Post-processing Model for Korean OCR using Synthesized Statistical Information

In this paper, we describe a contextual Korean OCR post-processing model considering unknown words. This work starts from the following premises: 1) In the language having very large character set, it is hard to directly correct erroneous string; 2) word formation is deeply related not only to morphological feature but also to phonological feature(esp. syllable combination on the surface level); 3) OCR post-processing system must rely on lexical and contextual information to correct OCR errors accurately. Based on the premises, the proposed system is composed of following three modules. First, it generates candidate words and evaluates the conndence of each candidate. Second, it selects feasible candidates by evaluating the word possibility of each candidate with its corresponding syllable connectivity and morphotactic connectivity. Third, it analyzes the contextual association among words and selects the most feasible word by the statistical information synthesizing the conndence weight, word possibility and contextual association strength of each candidate. Experimental results show that the proposed system can improve the performance of an OCR system from 94.1 to 97.6% by character-unit and from 87.6 to 97.1% by word-unit. In addition, this system can recognize 87.4% of unknown words.