Course contentsShow
← Machine Learning and Deep Learning
Lesson 1308 of 3,53829. Core NLP TasksFree lesson

Conversational Question Answering

Handling multi-turn dialogues where questions depend on previous context and conversation history.

Conversational Question Answering

What You'll Learn

You'll learn how question answering systems handle multi-turn conversations where later questions depend on earlier context and dialogue history.

The Challenge of Context-Dependent Questions

Unlike standalone question answering (where each question is independent), conversational QA requires tracking dialogue history across multiple turns. When someone asks "Who founded Microsoft?" followed by "When did he retire?", that second question only makes sense with the first question's answer (Bill Gates) in mind.

Think of it like a conversation with a knowledgeable friend who remembers what you've been discussing—not someone with amnesia who forgets each exchange immediately.

How It Works

Conversational QA systems extend the extractive and open-domain QA approaches you've learned by:

  1. Maintaining conversation history: Storing previous questions and answers as context
  2. Coreference resolution: Understanding pronouns ("he," "it," "they") refer back to entities mentioned earlier
  3. Context-aware encoding: Feeding both the current question AND conversation history to the model
  4. Dialogue state tracking: Keeping track of what's been discussed to resolve ambiguous references

The model typically concatenates previous turns with the current question when encoding, allowing the transformer to attend to relevant historical information when predicting the answer span.

Key Differences from Single-Turn QA

  • Input structure: [Previous Q1] [Previous A1] [Previous Q2] [Previous A2] [Current Question] [Passage]
  • Longer context windows: Must fit conversation history plus passage
  • Harder evaluation: Must handle pronouns, ellipsis ("And the capital?"), and implicit references

Key Takeaway: Conversational QA extends single-turn question answering by incorporating dialogue history, enabling models to resolve context-dependent questions through attention over previous turns and maintaining conversational state.