I have been having trouble getting chatGPT to transcribe a large and complex equation. What’s frustrating is that seems to make minor mistakes in some of the simpler terms in the equation. For example here’s the scan of the equation (it’s the one starting with p = … ):
chatGPT (4o) comes back with a denominator that missing the phi term, and many mistakes here and there.
On the other hand, if I pass along just a single line of the equation, it get’s the result exactly correct.
Maybe I should work out a way to subdivide the image and then feed it in one line at a time. But that seems like something the system should be doing on it’s own if it ends up producing more accurate transcriptions in general.
Is this a known problem? Am I maybe missing something and should be doing things differently?
I wouldn’t exactly say a problem and more of this is par for the course.
Yes, make sure the image is taken perfectly flat. Personally I use my copier/scanner to do such with the book firmly placed on the glass and pressed down so the entire part with the equations and about an inch around those is flat on the glass.
If you are trying for 100% recognition of such equations don’t waste your time trying to achieve that accuracy. If you can consistently get about 95% accuracy then do the rest manually consider yourself doing good.
While I have not tried using the latest AI often with such, try giving it the context before scanning and ask it to proof read the equations after being converted to see if it can fix up the equations before you have to do the manual fixes.
Let us know if you achieve success, I am sure others are doing similar to what you are doing and have similar problems and would like to know what helped you get better results.
