Question Answering (QA) systems automatically answer natural language questions, either by extracting answers from provided documents (extractive QA) or generating answers from learned knowledge (generative QA). Extractive QA, exemplified by SQuAD-trained models, identifies the span in a given passage that answers the question: 'When was the Eiffel Tower built?' answered by highlighting '1889' in the context. The model predicts start and end positions of the answer span within the document. Generative QA produces free-form answers without requiring supporting documents, relying on knowledge encoded in model parameters during pretraining. Large language models excel at this but can hallucinate plausible-sounding but incorrect answers. Open-domain QA combines retrieval and reading: first find relevant documents from a large corpus, then extract or generate answers from retrieved content. This RAG approach grounds answers in evidence. Multi-hop QA requires reasoning across multiple pieces of information. Conversational QA handles follow-up questions that reference previous context. Table QA answers questions about structured data. Visual QA answers questions about images. Evaluation typically uses exact match and F1 scores against reference answers.
Back to Glossary