questionanswering

Question answering is a field within artificial intelligence and information retrieval focused on building systems that provide direct answers to user questions. QA systems may operate over structured data such as tables or knowledge graphs, unstructured text such as documents and web pages, or multimodal sources. Tasks are often categorized as open-domain, where answers are drawn from a broad corpus, or closed-domain, which restricts the source to a specific collection. Systems can be extractive, identifying a span of text from a source, or abstractive, generating a new answer not present verbatim in the source. Conversational QA adds the constraint of maintaining context across multiple turns.

Typical approaches combine retrieval and reasoning components. In modern settings, retrieval-augmented methods use a retriever to fetch relevant passages and a reader or generator to produce a final answer. End-to-end neural models, especially those based on transformers, have become prominent, enabling direct answer generation or span extraction from retrieved material. QA tasks also include fact-checking, justification of answers, and multi-hop reasoning, where the correct answer requires combining information from several sources.

Common benchmarks include datasets such as large open-domain collections and reading comprehension corpora that test answer accuracy, completeness, and justification. Evaluation typically uses exact-match and F1 scores for factual answers, with additional metrics for readability and where appropriate, multi-hop reasoning and confidence indicators.

Applications span search engines, virtual assistants, customer support, education, and data analysis. Ongoing challenges include handling ambiguous or under-specified questions, minimizing reliance on any single source, improving robustness across languages, and ensuring factual accuracy and transparency.