An Efficient OCR Error Correction Method for Japanese Text Recognition

OCR error correction using Japanese morphological analysis contains two time-consuming procedures: extraction of candidate words from combinations of candidate characters, and finding the most plausible word sequence in combinations of the candidate words. In this paper an optimal word extraction technique, and the use of lexical entries that are tailored for Japanese verb inflection, are investigated and developed. Compared to a standard method, the new method requires 84% less computation, and captures 2.6% more candidate words. The new design of lexical entries reduces the chart parsing computation by 20%. The error correction rate of the system is 86.9%, which is 19.6% higher than that of the standard one. 1 Introduction In ordinary Japanese written sentences, words are not separated by spaces. Thus error correction algorithms for Japanese text recognition using Japanese morphological analysis is forced to use two time-consuming but indispensable procedures: (1) Searching for lexical entries in combinations of characters arranged in a "candidate character lattice" (a sequence of lists of candidate characters output by a character recognition module) and extracting candidate words from a dictionary at each position in the output character string, (2) Finding the most plausible word sequence from the "candidate words lattice" (the set of words extracted by procedure (I)). Those procedures take most of the error correction time. Procedure (1) consumes more than half of the processing time [I]. Thus, a more efficient word extraction method would be helpful. At the same time, word extraction must be sufficiently precise so that as many words as possible are found. This is due to missing words resulting in linguistic post processing failure, which consequently results in a lower error correction rate. This paper investigates an optimal word extraction technique and the use of lexical entries specially tailored for Japanese verb inflection. Compared to the standard method Extl(cl) (mentioned in section 2). this method requires significantly less computation and attains a higher error correction rate even if the recognition rate of the first candidate character is quite low. The comparative results of experiments will also be shown.