Train a printed-Hebrew Kraken model, according to the instructions here. The goal is to quickly test Kraken on Hebrew (i.e. not HTR of a manuscript). Since Kraken requires at least 800 lines transcribed text, and the process provided by Kraken is manual, an automation script was developed, to run all the preparations of the lines image-text pairs. After the training set is ready, the Kraken model training can be run.

Source text

The training text used is Genesis book, taken from Sefaria. The text is in Creative Commons license, taken from (http://www.tanach.us/Tanach.xml)

Requirements

On a Debian Linux (Ubuntu 16.04 works well), install Kraken along with the Python installation and Vagrant image required for the training. Instruction here. The script requires:

Beautiful Soup 4 Python module:

pip install beautifulsoup4

ImageMagic to convert text into images:

sudo apt-get install imagemagick

Training set preparation

Run the script src/train_kraken_print.py The script does the following:

• Breaks the Genesis JSON file into chapters, and for each chapter:
  • Extracts the verses into text file, verse per line.
  • Generates an image, big enough for Kraken - 5000x8000px, using ImageMagic. Since ImageMagic doesn't seem to support LTR, the Hebrew words are reversed. It's simplistic, but works for pure-Hebrew text. For mixed text, bidirectional support is required.
  • Runs the Kraken ketos utility, to generate the HTML file, intended for manual transcription
  • Fills the HTML transcription file automatically, with the text from the matching text file.

Running the script generates HTML files of the transcribed images to ./work/ketos_filled_*.html, along with byproducts.

Extract lines image-text pairs

Now that the transcription file are ready, run the Kraken command: ketos extract --reorder --output training --normalization NFD work/ketos_filled_* The files will be generated into training/ directory.

Kraken training

Get and run the Kraken Vagrant virtual machine. Make sure to run Vagrant in the same directory in which the train_kraken_script.py script was run: vagrant init openphilology/kraken vagrant up vagrant ssh In the Kraken VM, run the training command: ./train.sh /vagrant/training heb_print It will take the training set, and train a model called heb_print.