질문 답변

Question answering returns an answer to a question under a defined task and evidence setting. Extractive reading-comprehension systems select a text span from a supplied passage. Generative systems produce answer text rather than require an exact source span, although generated wording can still overlap a source. Another distinction concerns evidence access. Closed-book QA relies on model parameters without an external document lookup at answering time. An open-book system can use supplied or retrieved documents. Retrieval-augmented generation combines a retrieval component with generation, providing passages that can support an answer. Retrieved text is not automatically correct, relevant, or faithfully used, so that architecture does not guarantee factuality. The cited RAG research compares particular systems and tasks, rather than proving a universal hallucination reduction. Extractive models commonly assign scores to possible answer starts and ends. Decoding selects a suitable span subject to boundary constraints; independently high boundary scores should not imply that an end before the start is a valid answer. Some questions have no answer in the given evidence. SQuAD 2.0 was designed to include unanswerable questions, making answerability detection important rather than rewarding a plausible-looking span for every question. Evaluation has several dimensions. SQuAD-style exact match checks the answer against reference wording after the evaluation's normalization. Its token-overlap F1 gives partial credit for overlapping answer words, and does not measure every form of factual correctness. Testing paraphrased questions adds another dimension: the same intended question can use different wording. Research on paraphrased SQuAD questions found performance changes for evaluated models, showing why a score on the original wording alone can conceal sensitivity.