Course contentsShow
← Machine Learning and Deep Learning
Lesson 1304 of 3,53829. Core NLP TasksFree lesson

Abstractive Question Answering

Generating answers in natural language rather than extracting spans, using encoder-decoder models.

Abstractive Question Answering

What You'll Learn

You'll discover how models can generate natural language answers from scratch instead of just copying text spans from a passage.

The Core Idea

Until now, you've seen extractive QA, where models highlight existing text snippets as answers (like BERT finding spans in a passage). Abstractive QA takes a different approach: the model generates answers in its own words, synthesizing information and potentially paraphrasing or summarizing.

Think of it like the difference between a student who underlines sentences in a textbook (extractive) versus one who reads the material and explains it in their own words (abstractive).

How It Works

Abstractive QA relies on encoder-decoder architectures you've already learned about (like T5 and BART):

  • The encoder processes both the question and the context passage
  • The decoder generates the answer token-by-token, word by word
  • The model can reformulate information, combine facts from multiple sentences, or even reason about implicit details

For example:

  • Question: "Why did the expedition fail?"
  • Context: "The team ran out of supplies on day 12. They had to turn back before reaching the summit."
  • Extractive answer: "ran out of supplies" (copied span)
  • Abstractive answer: "The expedition failed because supplies were depleted before they could reach their destination." (generated synthesis)

Advantages and Challenges

Abstractive QA can handle questions requiring synthesis, paraphrasing, or multi-hop reasoning more naturally. However, it's more prone to generating factually incorrect "hallucinations" since it's not constrained to existing text.

Key Takeaway: Abstractive Question Answering uses encoder-decoder models to generate natural language answers rather than extract spans, enabling more flexible and human-like responses at the cost of potential factual errors.