Background
https://readcoop.eu/transkribus/ is a “comprehensive platform for the digitization, AI-powered text recognition, transcription and searching of historical documents.” I’ve heard it mentioned repeatedly at recent conferences, and I’m intrigued by the apparent possibilities. Could we make use of it, in the next few years, to make transcribing handwritten documents cheaper and more efficient?
I’m proposing to spend a bit of time tinkering with it in order to find out.
Table of contents
Scope and assumptions
High quality translation of text files (as opposed to images of handwriting) is likely to become cheap, easy and convenient in the next 10 years or so, due to progress in machine learning and the economic incentives involved. This will be especially true for translations between the major modern languages.
I am focusing strictly on transcription here, not on translation.
An expert human transcriber and translator, given enough time, is always going to do a better job than an automatic transcription engine, because the expert can infer the best reading of a handwritten text from deep familiarity with the linguistic, historical and cultural context within which a document was produced.
Experts’ time and effort are precious commodities; we don’t want an expert to waste any time on mere data-entry.
It’s at least possible to imagine that a decent machine transcription of an image into a text file might be a real time-saver to an expert, even if the copy includes a certain number of errors.
Timeline
roughly 3 weeks in mid-late January 2023 (subject to interruptions, of course; in the short term, this ranks low among our priorities).
General approach
Get familiar with the Transkribus user interface
Learn the lingo (and there is a lot of lingo)
Explore the user community and other online resources
Research current use patterns: what are common ways in which our peers are using the tool
Explore how an institution like us might make use of the tool
Tasks
Formulate proposal
Install Transkribus and create an account that I can use
Read the docs and watch the videos and go through the tutorials
Test the tool using Jocelyn's transcriptions of Bredig handwritten letters as "ground truth"
Attempt to train the software to transcribe letters it hasn't seen yet
Evaluate automatic transcriptions results against expert transcriptions
Take a look at other collections of handwritten letters too (Pasteur? Booth?)
Write everything up (see below)
Deliverables
I’d like to write up my results here in the wiki; I want to produce a short, non-technical report containing:
a description of what the software is, and what it is not
a glossary of machine-learning technical terms
a list of online resources (websites, listservs, videos, etc.)
examples of how peer institutions have used the software
practical recipes that could get us more mileage out of human experts in the future, by partially automating the transcription process.