Page MenuHomePhabricator

Add functionality to utilise line detection models on Wikisource
Closed, ResolvedPublic

Description

Give users on Wikisource the option to utilise a line detection model while using the Transkribus OCR

  • Only if a line detection model ID exists for that particular wiki, show the option
  • Select whether or not to use the line detection model

Use changes made to the API as in T340837

Event Timeline

Restricted Application added a subscriber: Aklapper. · View Herald Transcript
PMenon-WMF changed the task status from Open to In Progress.Aug 10 2023, 9:46 AM
PMenon-WMF claimed this task.
PMenon-WMF removed a project: Community-Tech.

Change 947796 had a related patch set uploaded (by Parthiv; author: Parthiv):

[mediawiki/extensions/Wikisource@master] Add functionality to use line detection model

https://gerrit.wikimedia.org/r/947796

Change 947796 merged by jenkins-bot:

[mediawiki/extensions/Wikisource@master] Add functionality to use line detection model

https://gerrit.wikimedia.org/r/947796

Merged.

But I have a couple of questions:

Under what situations would someone want to not send the line model? It seems to me that it should at least default to sending it, and perhaps not even be a user-facing option to not send it. (In the on-wiki UI this is, not in the advanced form.)

Is it clear to users what a line detection model is and why they'd want to change the setting?

So by default, Transkribus uses a line detection model called Mixed Line Orientation (MLO). A user on, say, Balinese Wikisource has two options

  • Choose not to use line detection model on wiki => Transkribus uses MLO in the background
  • Choose to use line detection model on wiki => Transkribus uses the line detection model configured on that Wikisource (a custom line detection model for that language/writing style)

The thing with providing users a checkbox is because it gives users the flexibility of using either the MLO model or the language specific model. This is important as both the models might give different results in the output text depending on the suitability of the line detection model to the image being processed.

I suppose the only situation where a user would not send a line detection model is when they don't know what it is or what it does! Maybe we should work on implementing a one time pop-up (like the blinking indicator when the page first loads) that explains the necessity of line detection models clearly.