Following the first few steps on the Google Code wiki for Tesseract to learn how to train it for a new language, I've used the moshpytt box editor on a sample file, and read through the other sample data. It looks like we may be able to do something like the following, For any sufficiently large run of a journal which has consistent page-images, fonts, print quality etc.:

• Create an imagemagick script which optimizes the images for OCR.
• OCR some sample poems and use moshpytt to correct the results.
• Go through the rest of the training process, to create a complete training set for "Victorian English".
• Use a standard dictionary, but tweak it to remove any modernity that's distracting, etc.
• Add XSLT for markup to the end of the toolchain.
• Add a step to pull in metadata from the db.
• Run it on the whole set, and get decent TEI-encoded transcriptions out the other end.