I’m trying to take long form, flowery writing and extract concrete things like hair color.
I’m not sure whether to go down an examples route, with Classification, or more free-form natural language and asking “what is their hair color?”
Classification seems like it will be laborious to generate examples. (Flaxen means blonde, dark means black, auburn means brown…)
But a more natural language completion like “what is their hair color?” gives more unpredictable results like “her hair color is black” rather than just “black”.
So for extracting discrete information, is Classification better, and I just need to get better at making the example data?
With GPT-3 it’s neither. The term extraction or distillation would be better. This task is most similar to NER or keyword extraction (both of which are NLP).
some hyper parameters to tune-
Max_tokens: set to 1 or 2
Give an example (or two) of the question answer pair after the instruction.
Answer the question in one word based on the text below, similar to examples below.
Example Question: What is the hair colour?
Example Answer: Black
A lot of these tricks went a long way when I was developing my product BookMapp
And providing examples gets me a lot farther, thank you.
I am used to the rigidity of formal code, so a task that needs to be convinced or guided or nurtured, is such an orthogonal concept to me.
BookMapp sounds like high-level stuff… an example on the page would be really helpful. Appreciate there are credits gratis on signup
Thanks for your kind words on BookMapp. Would love to hear what you think, email me:
contact at BookMapp dot com
Yes, it does take time to get your head around.
A tip, treat it like an idiot savant, high competence, low common sense. Deal with it like a brilliant intern.