It seems that nowadays lots of people building tools to extract data from different sources (e.g. PDF reports) with the help of models like GPT-4, LLama or Falcon.

I was wondering, if this really works, because from what I know, you could train GenAI models to reproduce entities in the same way as they are written in the original text, but still produce the output themselves instead of pointing to the parts of the text.

So I was wondering if Extraction can be achieved reliably with GenAI or if ExtractiveAI with Bert-like models (e.g. SpanMarker for NER) is the way to go?

  • $\begingroup$ just because they can, doesn't mean that they are as good... and even if they are, running GPT-4 with respect to a Bert-like specialized model, is some order of magnitudes more expensive $\endgroup$
    – Alberto
    Sep 14, 2023 at 11:11
  • $\begingroup$ i have the same question. I would like to build a q&a system but it should be reliable. Extractive q&a systems extract answers from the provided context. That means you have no hallucination. You could have a wrong extracted answer, but the answer will be always from existing content and not generated. While generated answers could be totally wrong and seem like correct. As far as I understand, its easier to build a q&a system with generative models like GPT4 because they have a crazy capability to also extract the answer but it could still be wrong (using RAG). $\endgroup$
    – Khan
    Dec 2, 2023 at 17:26
  • $\begingroup$ on the other hand using extractive q&a you need training data to adapt the model to your specific domain. That requires more work. But as I said, I am also thinking about all this stuff since I am relatively new in this area $\endgroup$
    – Khan
    Dec 2, 2023 at 17:31


You must log in to answer this question.

Browse other questions tagged .