Question Answering Task Overview
What you'll learn: You'll understand the main types of question answering systems and how they differ in approach and scope.
What Is Question Answering?
Question Answering (QA) is an NLP task where a model receives a question and must provide an answer—either by extracting it from a given context, generating it from scratch, or retrieving it from a knowledge source. Think of it like the difference between finding a quote in a book (extractive) versus explaining something in your own words (abstractive).
Four Key QA Task Types
Extractive QA
The model identifies and extracts a span of text directly from a given context passage. The answer already exists word-for-word in the input.
Example: Given a paragraph about Abraham Lincoln and the question "When was Lincoln born?", the model extracts "February 12, 1809" from the text.
Abstractive QA
The model generates an answer in its own words, possibly summarizing or paraphrasing information. The exact answer phrase may not appear in the source.
Example: Answering "Why did Lincoln issue the Emancipation Proclamation?" might require synthesizing multiple facts into a coherent explanation.
Closed-Domain QA
The system operates within a specific topic or dataset. Questions and answers are constrained to a narrow knowledge area (e.g., medical records, company documents).
Open-Domain QA
The model can answer questions about any topic, often requiring retrieval from large knowledge bases or the entire internet. This is much harder—like being asked anything versus being quizzed only on one textbook chapter.
Why This Matters
Each QA type requires different architectures and approaches. Extractive QA can use token classification (similar to NER), while abstractive QA needs generation capabilities (like T5 or GPT). Understanding these distinctions helps you choose the right model and evaluation strategy.
Key Takeaway: Question answering splits into extractive (selecting text spans), abstractive (generating answers), closed-domain (narrow topics), and open-domain (any topic) systems—each requiring distinct modeling approaches.