If the concern with using generative models for question answering is that these models aren't always producing factual information, why is it that people are using these models with Retrieval Augmented Generation (Open Generative QA/Closed Generative QA), rather than using a transformers-based extractive QA model that can refer users to the potential answer from actual text documentation?


1 Answer 1


The issue is that extractive QA limits the types of questions to ones that have answers that appear exactly in the text.

For example, consider the question:

Is Bob considered a good movie director?

It's very unlikely that you're going to have a document that says exactly Bob is a good movie director. But, you might have several documents that talk about the quality of movies that Bob has directed or interviews that discuss how well he works with actors etc. from which you can infer an answer from the totality of the evidence.

That's not say that extractive systems are always worse, you're just faced with the classic trade-off of a restrictive but predictable system versus a more flexible system that may e.g., hallucinate.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .