ChatGPT has become my go-to LLM for transcribing printed documents. I have many, many pages of typed text from my father's writings that I want to digitize. For a few months now I have been using this Free Online OCR image-to-text converter, and it has served me well for printed pages. It's far from perfect and requires proofreading and a fair amount of correction, but I found it far, far better than any other such tool over the years. And for the price, I could live with the five-pages-a-day limit. I was content.
Then, on a whim, I decided to see what ChatGPT could do. I uploaded one of my pages, and was blown away. Nearly perfect copy, in just a few seconds.
I still proofread everything. Not only for quality assurance, but because I'm interested in the stories. It took a little while for the LLM and I to work out just what I wanted, but now it knows to give me a faithful transcription (which turns out to be more accurate than what I had been using) along with some "light editing." It knows I want my father's writing to stay intact, but it corrects spelling errors and typos, and flags places where it finds the text ambiguous possibly in error. For example, my father wrote about hearing a lecture by Galen Rowen. ChatGPT suggested that the correct name might be Galen Rowell. A little research convinced me that the LLM was correct.
At first, it annoyed me by "fixing" the paragraph breaks—my father liked long paragraphs, and Chat GPT decided they would be better broken up. But on proofreading, I often decided that the story did flow much better with more paragraphs, so I decided to let it do its thing on that and manually fix anything I thought should go back together. I did stop it from changing words my father had put in all caps into lower case. That we tweaked by instructing it to leave all caps as is if the word was three or more characters long. This allowed it to continue to correct words in which the capitalization had been incorrectly typed to carry over to the next letter. The adjustments take time to figure out, but the hours saved in the end are priceless.
Similarly, ChatGPT is invaluable when I'm extracting text from e-mails which are formatted so that there's a paragraph mark at the end of every line. Or written so that every (real) new paragraph begins with a string of spaces. (Remember, I'm working with thirty-year-old text.) It turns out that ChatGPT handles that very well, and applies the same standards we developed with the OCR work.
Heretofore, most of my time playing with AI has been just that: playing. But this has the potential to do as much for my productivity as replacing logarithm tables with a calculator.
