Introduction to eScriptorium, HTR for Hebrew Manuscripts, part 2

The first part of this post covered starting a project with eScriptorium and properly segmenting a page to prepare it for automated transcription. It commended a combination of automated and manual steps, beginning with an automated segmentation that I manually cleaned up and created new regions for. The whole process took only a few minutes for a page.

Having finished the segmented the folio, I returned to the “images” tab and selected the folio I had just been editing. There, I clicked the “transcribe” button all the way on the right side of the screen. It asked me to select a model and I chose the appropriate Hebrew model.

Preparing for Automated Transcription

The algorithm ran over the image and deciphered the text. The whole process took less than a minute for a single folio (again, with music also streaming in the background). Now came the moment of truth: time to check the results. I clicked on the edit button on the folio I had just run through the transcription. In order to have a better overview of what it recognized, I clicked on the second button from the top right, “Transcription”. It opens a display of the recognized text next to the image. (The right-most button displays the text in reading order, not as it appears on the image.)

The Image and Its Automatically Transcribed Text

At first glance, the results look very impressive. A closer look confirms that, but there are some errors that crept in to the transcription. None of these were unanticipated.

To really authenticate the results, I clicked on the first line of the displayed transcription (not on the image of the folio). Clicking the transcription that way brings up an image of the segmented line as displayed by the mask. Here, you can read the line of the image precisely and check the transcription against it.

An Error in the Automatic Transcription

As the image of the first line with transcription demonstrates, there was an error in the transcription. Fixing it was not a problem. I changed my keyboard setting to Hebrew and corrected the errant letter. Having finished with the first line, I used the down arrow to proceed to the second line. It had no errors. Thus I proceeded through the whole first column. The number of errors was only about one per every two lines. The most common error was spacing, understandable since the spacing between words is not always great in Masoretic manuscripts. Other errors were precisely those that humans make as well: confusing ב and כ; ז, ו, and ן; י and ו; עי and ש; etc. The most common errors stemmed from my mistakes, particularly cases in which my lines failed to include either the beginning or ending of the line. The model incorrectly identified the letters since a piece was missing, but it satisfactorily recognized portions of letters as similar looking letters.

Checking the transcription of a single column (28 lines) and correcting the errors took less than ten minutes. The whole process was straight-forward and generally user-friendly. That translates into only a few hours work for the whole portion of the manuscript that is relevant to my work. Quite a time saver. That being said, there are some issues with eScriptorium for biblical manuscripts.

First and foremost, it conspicuously does not currently recognize vocalization and cantillation. That means that these elements—where necessary—must be added secondarily. That would be quite time consuming. I was told that this will be emended and that it is possible to train your own model to do this based on the models they have. That would require some substantial input, but would presumably be easy enough to generally automate it over time for large projects. For a few folios, it’s probably not worth it.

As far as I can tell, there is also no way to manually adjust the reading order of the lines. Other OCR products, like OCR4All do have such an option. That being said, it does not have Hebrew compatibility without developing your own models. (At least it did not the last time that I worked with it.) Those present my primary issues with it. Particularly if you are working with unvocalized manuscripts, there are likely few issues with eScriptorium, though it will certainly require a little practice.

If you do start to work with the application, I have a few tips to hopefully speed your workflow:

  • The interface works quickest and easiest if you work ambidextrously, i.e., with one hand manipulating the mouse and the other hand working the keyboard.
  • That means that you need to learn the most important keyboard shortcuts, which are easy enough to learn and generally stem from English terms for the processes they include.
  • Don’t forget to use the escape key. If you ever click anyway and start drawing a line or region, don’t finish it and then delete it; just click escape.
  • The more time you spend segmenting, the better your results will be.
  • Don’t cut your lines too short, whether at the opening or the closing. Lengthening the lines is easy, even several at a time. Make sure to check them well.
  • Have fun!

I hope that this introduction has been helpful for you. I can recommend eScriptorium to largely automate the process of transcribing biblical Hebrew manuscripts, even with the caveats noted above. If you have other tips and tricks, I would love to hear them.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s