Stop 09 · Natural Language ProcessingTopics · 10/10

Question Answering Models

In short

Ask a question, get the exact answer from a document. QA models power help-desk bots and document search, and they're a key building block behind retrieval-augmented chatbots.

Models that extract or generate an answer to a question given a context passage — the core mechanism behind search-augmented and RAG-based assistants.

The blob behind a railway enquiry window runs a finger down a big timetable board without any writing.

Think of it like this

The railway enquiry counter

You ask the enquiry clerk when the evening train to Chennai leaves. He runs his finger down the timetable, stops at the right row and points: 6:40 pm. He doesn't invent anything, he finds it. An extractive QA model points to the answer inside your text.

Watch it happen

Finding the answer span

Library

opens

and

closes

1/4Question: 'When does the library close?' The model reads the question and this context together.

The big ideas

What you need to know

Extractive QA predicts a span

The model outputs a start score and an end score for each token, then picks the best valid span.

Sometimes there's no answer

Models trained on SQuAD 2.0 can say 'not in the text' instead of forcing a guess. Check the score.

Retrieve first, then read

For many documents, first find the few relevant passages with search, then run QA on those. That's the core of RAG.

Generative QA

LLMs write answers in their own words. More flexible, but they can make things up if the context lacks the answer.

See it in code

Walk through the code

python

1from transformers import pipeline2 3qa = pipeline("question-answering", model="deepset/roberta-base-squad2")4 5context = ("The central library opens at 8 am and closes at 10 pm on weekdays. "6           "On Sundays it is open only from 10 am to 4 pm.")7 8for q in ["When does the library close on weekdays?", "Is it open on Sundays?"]:9    ans = qa(question=q, context=context)10    print(q, "->", ans["answer"], round(ans["score"], 2))

1/3Load a model trained to find answer spans, including questions that have no answer.

A RoBERTa model fine-tuned on SQuAD 2.0. The answer is always a span copied from the context.

Save yourself the pain

Common mistakes

Oops

Expecting an extractive QA model to answer yes/no or summarise reasons.

Do this instead

It can only copy a span. For yes/no or explanations, use a generative model with the context.

Oops

Passing a whole book as the context.

Do this instead

Contexts longer than the model's limit are split into windows. Retrieve the relevant passages first for speed and accuracy.

Your mission

Build a mini FAQ bot

~30 min

Quick check

What does an extractive QA model actually output?

Go deeper

1Run a pretrained extractive QA model (e.g. fine-tuned BERT on SQuAD) on real passages.
2Learn the difference between extractive QA and generative QA.
3Understand how this connects directly to the retrieval-augmented generation you'll build later.