Extraktive und generative QA

Question answering, or QA, is the task of returning a precise answer to a natural-language question instead of a list of links. You ask, the system answers. 💬

QA systems split into two families. Extractive QA pulls the answer straight out of a given passage, word for word.

The other family is generative QA. Instead of copying text, the model writes a fresh answer in its own words, possibly combining several facts.

Extractive QA always works against a context passage you provide. The answer must be a span that already exists inside that text.

In extractive QA the output is a span: a start and end position in the context. Here the span is simply the word Warsaw.

A generative model is not limited to the passage. It can summarize, rephrase, or answer from knowledge baked in during training. ✨

Extractive answers are easy to trust: you can point to the exact text they came from. Generative answers are flexible but harder to verify.

Because generative models invent wording, they can hallucinate facts that sound right but are false. Extractive QA cannot do that.

Choose extractive QA when answers live inside trusted documents and you need exact, auditable quotes, like legal or medical text.

Choose generative QA when you need fluent, combined answers or want to respond even when no single passage holds the full reply.

We begin with extractive QA because it is concrete and easy to evaluate. You will load a passage and pull the exact answer from it.

Test your grasp of the two QA families.

QA returns precise answers. Extractive copies an exact span from a context; generative writes new text. Extractive is auditable and where we start. 🎯