Post-Correction of Icelandic OCR Text
Jón Friðrik Daðason · 2012
The topic of this thesis is the post-correction of Icelandic OCR (optical character recognized) text. Two methods for spelling correction of OCR errors in Icelandic text are proposed and evaluated on misrecognized words in a digitization project which is ongoing in Alþingi (the Icelandic parliament). The first method is based on a noisy channel model. This method is applied to nonword errors, i.e., words which have been misrecognized during the OCR process and transformed into another word which is not in the Icelandic vocabulary. This method achieves a correction accuracy of 92.9% when applied to a test set of nonword errors from a large collection of digitized parliamentary speeches from the Alþingi digitization project (a total of 47 million running words from the years 19591988). The second method uses Winnow classifiers, and is applied to real-word errors, i.e., words which have been misrecognized during the OCR process and transformed into another word which also exists in the Icelandic vocabulary. A Winnow classifier is able to correct real-word errors by detecting words which do not fit in the context in which they appear and suggesting other similar words which are more likely to be correct. When applied to a test set of real-word errors from the same set of digitized texts as above, this method achieves a correction ratio of 78.4%. When both methods are applied to all errors in the digitized parliamentary speeches, an overall correction accuracy of 92.0% is achieved.