Multi-Hop Reasoning in QA
What you'll learn: How question answering systems can connect information across multiple sentences or passages to derive complex answers.
The Core Challenge
In simple extractive QA (like SQuAD), the answer usually sits in a single sentence within one passage. But real-world questions often require connecting dots across multiple pieces of information. Multi-hop reasoning means your model must "hop" from one fact to another—sometimes across different sentences or even different documents—to reach the final answer.
Why It's Different
Imagine answering: "What is the capital of the country where Marie Curie was born?"
Your reasoning chain:
- First hop: Find where Marie Curie was born → Poland
- Second hop: Find the capital of Poland → Warsaw
No single sentence contains "Marie Curie" and "Warsaw" together. The model must retrieve relevant facts, understand their relationships, and chain them logically.
How Models Approach This
Multi-hop QA systems typically:
- Use context encoding (from lesson 1301) to understand multiple passages
- Identify which facts are relevant to the question
- Build reasoning chains by connecting entities and relations
- Predict spans or generate answers based on aggregated evidence
Unlike handling unanswerable questions (lesson 1302) where evidence is missing, multi-hop scenarios have all the needed information—it's just scattered.
The Architecture Implications
Models must track:
- Which information they've already used
- What still needs to be found
- How facts relate to each other
This requires more sophisticated attention mechanisms and sometimes explicit reasoning modules beyond basic span prediction.
Key Takeaway: Multi-hop reasoning extends QA beyond single-sentence answers, requiring models to chain together facts from multiple locations to solve complex questions—mimicking human step-by-step logical reasoning.