3 ms·

An LLM? Just what I always wanted - an OCR tool that hallucinates.

spookie 1y ago
While dictionaries are often reliable enough to get the correct words back when some glyphs are misrecognised I wonder if some type of LLM would help in some cases. Not a worry in modern digital first documents though.
spwa4 1y ago
You don't? Think about it. If your picture/source data is not perfectly clear ... what do you want? We all want perfection, but if you can't have that ... Would you prefer what current OCR does and just suddenly sentences go 2#!@%7Q&*@3 ladfk !@$? Or would you rather have a reasonable completion of a sentence that is nearly always (but not quite always) correct, that even actually takes the context into account?
- duskwuff 1y ago
  > Would you prefer what current OCR does and just suddenly sentences go 2#!@%7Q&*@3 ladfk !@$? Yes, actually. I'd rather be aware that the OCR tool failed somewhere than have the tool silently fabricate part of the text, or "correct" perceived errors which were present in the source document.
  - boredhedgehog 1y ago
    But you aren't aware, because the OCR doesn't know that it failed. You would have to go through the entire text by hand to fix the corruptions, but that's too much work, so you won't, and the corruptions stay in. In practice and at scale, the guesses of the LLM are the superior outcome.
    - thaumasiotes 1y ago
      > But you aren't aware, because the OCR doesn't know that it failed. You would have to go through the entire text by hand to fix the corruptions, but that's too much work, so you won't, and the corruptions stay in. Well, if you assume that you're never going to read the book, then sure. But in that case it's even more efficient to not OCR the book either. You'll never know the difference. If you do read the book, you'll know where the failures are. And they're easy to correct if you can edit the document. I usually file reports of printing errors in Kindle books when I encounter them. (Do the errors get corrected? No.)
- akho 1y ago
  Your picture of a ebook is perfectly clear.

While dictionaries are often reliable enough to get the correct words back when some glyphs are misrecognised I wonder if some type of LLM would help in some cases. Not a worry in modern digital first documents though.

spwa4 1y ago

You don't? Think about it. If your picture/source data is not perfectly clear ... what do you want? We all want perfection, but if you can't have that ... Would you prefer what current OCR does and just suddenly sentences go 2#!@%7Q&*@3 ladfk !@$? Or would you rather have a reasonable completion of a sentence that is nearly always (but not quite always) correct, that even actually takes the context into account?

duskwuff 1y ago
> Would you prefer what current OCR does and just suddenly sentences go 2#!@%7Q&*@3 ladfk !@$? Yes, actually. I'd rather be aware that the OCR tool failed somewhere than have the tool silently fabricate part of the text, or "correct" perceived errors which were present in the source document.
- boredhedgehog 1y ago
  But you aren't aware, because the OCR doesn't know that it failed. You would have to go through the entire text by hand to fix the corruptions, but that's too much work, so you won't, and the corruptions stay in. In practice and at scale, the guesses of the LLM are the superior outcome.
  - thaumasiotes 1y ago
    > But you aren't aware, because the OCR doesn't know that it failed. You would have to go through the entire text by hand to fix the corruptions, but that's too much work, so you won't, and the corruptions stay in. Well, if you assume that you're never going to read the book, then sure. But in that case it's even more efficient to not OCR the book either. You'll never know the difference. If you do read the book, you'll know where the failures are. And they're easy to correct if you can edit the document. I usually file reports of printing errors in Kindle books when I encounter them. (Do the errors get corrected? No.)
akho 1y ago
Your picture of a ebook is perfectly clear.