What is Question Answering

How It Works

Extractive QA identifies the answer span within a given context passage by predicting start and end token positions. Abstractive QA generates the answer in natural language, potentially synthesizing information from multiple sources. Open-domain QA first retrieves relevant documents from a large corpus, then extracts or generates answers from the retrieved context.

Technical Details

Extractive models fine-tune BERT or RoBERTa on SQuAD-style datasets to predict answer spans. Generative QA uses encoder-decoder models (T5, BART) or LLMs with RAG (Retrieval-Augmented Generation). Multi-hop QA requires reasoning across multiple documents. Evaluation uses Exact Match (EM) and F1 token overlap for extractive, and ROUGE/human evaluation for generative answers.

Best Practices

• Use RAG-based QA to ground answers in retrieved documents and reduce hallucination
• Implement answer confidence scoring and return 'no answer found' for low-confidence responses
• Provide source citations alongside answers for user trust and verification
• Fine-tune QA models on domain-specific question-answer pairs for specialized applications

Common Pitfalls

• Returning plausible-sounding but incorrect answers without confidence thresholds
• Not handling unanswerable questions where the context does not contain the answer
• Using extractive QA for questions requiring multi-document reasoning or synthesis
• Ignoring the quality of the retrieval step which directly determines answer accuracy

Advanced Tips

• Combine text QA with visual QA for answering questions about multimodal content
• Implement multi-hop QA with chain-of-thought reasoning for complex questions
• Use QA as an evaluation framework for testing retrieval pipeline quality
• Apply conversational QA with context tracking for multi-turn interactions over document collections