I'm working with a lot of scanned documents that I often feed into AI, like ChatGPT. Unfortunately, the output is frequently inaccurate because the OCR misreads the documents. How do you all typically detect or deal with problematic OCR before conducting any analysis? Do you prefer doing manual checks, or do you utilize any specific tools for this?
4 Answers
OCR technology can be hit or miss. If the accuracy of your data is crucial, it’s wise to have a human review it before proceeding.
You could also compare the text against a dictionary using the Levenshtein distance to identify potential errors. Names can be tricky, but it’s a good general approach.
I recommend trying Mistral OCR 3. No OCR solution is 100% reliable without some human oversight.
Your question touches on an important point. OCR isn't perfect—especially AI-based OCR. It's really about enhancing the speed of data processing while still ensuring validation is done by humans. One way to check for errors is to use different OCR tools and see if they produce the same results. If they do, there’s a good chance they interpreted the data correctly, but trained eyes are still needed for certainty.
That does help for obvious mistakes, but I'm concerned about subtle errors that look fine at first glance but can lead to significant issues later. Those are trickier to catch.