Question Answering

Question Answering systems answer natural language questions by extracting a span or generating a reply.

Extractive QA, as in SQuAD-trained models, finds the span in a passage. " is answered by highlighting "1889" in context. The model predicts start and end positions of that span. Generative QA writes a free-form answer without a required supporting document, using facts stored in parameters during pretraining. Large language models are strong here and can also hallucinate fluent wrong answers.

Open-domain QA retrieves then reads: find documents in a large corpus, then extract or generate from that evidence. That RAG pattern grounds answers. Multi-hop QA reasons across several facts. Conversational QA handles follow-ups that refer to earlier turns. Table QA answers questions about structured data. Visual QA answers questions about images.

Exact match and F1 against reference answers are the usual metrics.

Extractive QA is span prediction: start index and end index inside a passage, as in SQuAD, where "1889" is highlighted for the Eiffel Tower question. Generative QA writes tokens with no guarantee they appear in a document, so hallucination is the failure mode. Open-domain RAG first retrieves, then reads, so answers can cite evidence. Multi-hop items need two facts, not one sentence.

Conversational QA resolves "it" using prior turns. Table QA and visual QA swap the evidence type. Exact match and F1 are strict when wording differs from the reference. SQuAD (2016) asked models to span-extract answers from Wikipedia paragraphs. RAG later added retrieval in front of generation.