Usage of CorrectOCR is divided into several successive tasks.
To train the software, one must first create or obtain set of matching original uncorrected files with corresponding known-correct “gold” files. Additionally, a dictionary of the target language is needed.
The pairs of (original, gold) files are then used to train a HMM model that can then be used to generate k replacement candidates for each token (word) in a new given file. A number of heuristic decisions are configured based on whether a given token is found in the dictionary, are the candidates preferable to the original, etc.
Finally, the tokens that could not be corrected based on the heuristics
can be presented to annotators either via
or a HTTP
HTTP server. The annotators’
corrections are then incorporated in a corrected file.
When a corrected file is satisfactory, it can be moved or copied to the gold directory and in turn be used to tune the HMM further, thus improving the k-best candidates for subsequent files.