AI for OCR correction

[विश्वासो वासुकिजः (Vishvas Vasuki)]

Source: https://twitter.com/marquezxavier/status/1607615809274400771

Wish this existed for sanskrit!

--

--
Vishvas /विश्वासः

[Shreevatsa R]

Current state-of-the-art in OCR (e.g. what Google does in Drive/Vision) is already to incorporate a language model to generate reasonable output. (So the equivalent of "oECLARATION" in the English above is already not going to be generated.) 

So the equivalent of what ChatGPT is doing here would be useful in cases where OCR-ed text exists but the image itself is not available: because in the latter case it's always going to be better to throw away the existing poor OCR and just do OCR from scratch.

--
You received this message because you are subscribed to the Google Groups "sanskrit-programmers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sanskrit-program...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/CAFY6qgF6KqjA24KPWVVM-aa7czdv3Zyu19FoEUyNg38EOBRAGQ%40mail.gmail.com.

[suha...@gmail.com]

Such a tool would be a godsend for editing manuscripts, especially when only one manuscript is available, or readings are very corrupt. I'm imagining something like DeepMind's language model for restoring ancient Greek inscriptions (Nature 2022), but with the additional ability to emit text under metrical constraint. If anyone knows how the metrical constraint can be imposed in the ML model, I'm interested to hear about it.

(It might also be fun to build a निषेधाक्षरी app with it)

[Mārcis Gasūns]

emit text under metrical constraint - can you give an example, please?

[Suhas Mahesh]

"emit text under metrical constraint - can you give an example, please?"

If we want to do text-infilling on damaged manuscripts, for instance:

_khānām idam ābhisārikam asāv āpat-triyāmāñjaliḥ

The model should infill with duḥkhānām and not sukhānām, respecting the Śārdūlavikrīḍita template.

Also with a text correction model. 

Let's say a devanagari manuscript reads julita in the place of jvalita.

This happened because it was copied from Sharada, where ju/jva look similar.

tava virahe mama tanvi julanti sakalāny angāni (second half of āryā)

The suggested correction should be jvalanti (or even skhalanti), and not the metrically incorrect galanti, lalanti, lolanti etc.

You received this message because you are subscribed to a topic in the Google Groups "sanskrit-programmers" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/sanskrit-programmers/iYXN2IAvMzQ/unsubscribe.
To unsubscribe from this group and all its topics, send an email to sanskrit-program...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/sanskrit-programmers/e8ce9e06-2661-4c23-bf3a-604ab54eff9an%40googlegroups.com.

[Suhas Mahesh]

Correction: please read sakalāni gātrāṇi in place of sakalāny aṅgāni.