Question answering is the task of producing answers to questions posed in natural language. It appears in multiple-choice settings, where a system selects from provided options, and open-domain settings, where it answers without choices. section 1.6
The main variations differ in how they use context. Extractive question answering selects an answer from supplied text or a table. Open-book generative question answering uses supplied context to produce an answer in its own words. Closed-book generative question answering receives only the question and generates an answer from information learned during training. section 1.6
Question answering depends on several capabilities and exposes important limitations:
- The system must represent the question and its context, parse what is being asked, and generate or select a response. section 2.2
- Closed-book answers draw on patterns and information acquired from large text collections during pretraining, while open-book answers can use supplied supporting information. section 2.2
- Obscure information, underspecified questions, and uneven web coverage can produce incorrect or made-up answers; assumptions may also reflect the dominance of English and Western sources online. section 2.2
From a fact question to reading comprehension
A system might answer “Which politician won the Nobel Peace Prize in 2009?” or “What music did Beethoven compose?” Question answering also includes reading comprehension: the system receives several sentences or paragraphs and answers a specific question, sometimes in multiple-choice, dialogue, or abstractive formats. section 1.6
