Open-book generative question answering produces a free-text answer from context supplied with the question. Unlike extractive question answering, it does not copy an answer verbatim; unlike closed-book generation, it does not rely only on information learned during training. section 1.6
The approach suits open-domain questions, where no fixed answer choices are provided, and connects closely with reading comprehension. It still depends on representing and interpreting both the question and its context accurately. Missing, obscure, or ambiguous information can lead to an incorrect or made-up answer. section 1.6 section 2.2
Question-answering approaches differ by where the answer comes from:
- Extractive QA selects an answer from supplied text or a table. section 1.6
- Open-book generative QA uses supplied context to compose an answer in its own words. section 1.6
- Closed-book generative QA receives only a question and generates the most likely answer from training. section 1.6
A contextual question
A system may receive a passage and the question “Which politician won the Nobel Peace Prize in 2009?” It should use the passage to produce a natural-language answer rather than merely copy a span. section 1.6
