engleski

Croatian OCR Error Correction Using Character Confusions and Language Modelling

Manual correction of errors produced by optical character recognition (OCR) is a time-consuming task. This paper presents an automatic post- processing system that utilizes various methods for improving the OCR results of Croatian language texts. The system relies on knowledge of general characteristics of OCR errors, as well as language-specific knowledge. Used methods include character confusions, a character n-gram model, and word-splitting. A statistical language model is used for ranking the generated candidates depending on the sentential context. Experimental evaluation, performed on newspaper texts supplied by the Croatian News Agency, shows an error rate reduction of above 20%. These results amount to about 36% of the performance of manual correction.

Natural language processing; OCR; character confusions; character n-grams; word merge errors; language model; Croatian language

nije evidentirano