Radiology dictation error corrector prompt

Hi there,
My name is Reuben, I’m a radiology resident from Melbourne Australia.
I’ve built a “Dictation Error Corrector” web app using text-davinci-003 to address I problem I and many radiologists have.

As you can see in the example above, “enostoses” has been changed to “osteochondromas”. This is not correct, and unfortunately continues to recur in my experience using the app. There are a number of other similar mistakes that continue to be made (“enlocated” becomes “encumbered / encompassed / ensured”). Presumably, the model has not seen a great deal of these specific words in its training corpus.
My prompt is very simple, and I have found that if I specify “permitted words” in the prompt the resulting output is less reliable.
I feel as though this problem could be solved with fine-tuning, however davinci is so much worse than text-davinci-003 at a base level.
Any thoughts on how I might be able to solve this dilemma :thinking:

You can easily just do a basic substring search and replace on the completion text, correcting the errors.


Lol that is a great idea, obvious in hindsight :sweat_smile: Thanks @ruby_coder for adding some perspective!

1 Like

You are welcome @reubenschmidt


1 Like