What Is Question Answering (information Retrieval)?

Question answering (QA) in information retrieval is the task of automatically producing a direct, precise answer to a natural language question by locating and extracting relevant information from a text collection, knowledge base, or structured data source. Unlike traditional document retrieval, which returns a ranked list of documents that may or may not contain the needed fact, a question answering system identifies the specific passage, entity, or value that constitutes the answer. The field emerged from natural language processing, information extraction, and database search, and gained wide visibility through shared evaluation tracks at NIST's Text REtrieval Conferences (TREC) beginning in 1999. Modern question answering systems underpin virtual assistants, enterprise search platforms, and question-answering interfaces to scientific literature.

QA tasks span a wide range of difficulty. A factoid question such as "What is the boiling point of water?" requires locating a single numerical or named entity. A list question requires enumerating multiple entities. A definitional question requires synthesizing a multi-sentence description. Complex questions involving temporal reasoning, multi-hop inference, or quantitative comparison require the system to integrate information from several passages or sources, which places far greater demands on both retrieval precision and language understanding.

Factoid and Extractive Question Answering

The dominant paradigm in open-domain QA divides the task into a retrieval stage and a reading comprehension stage. The retriever, often based on sparse lexical matching with BM25 or dense vector search with learned embeddings, identifies a set of candidate passages from a large corpus. The reader then scans these passages to locate the span of text that most directly answers the question, a process called extractive reading comprehension. The NIST TREC question answering track papers from the early evaluation periods document the transition from keyword matching to linguistically informed extraction methods and the metrics used to evaluate answer correctness. Stanford's SQuAD benchmark later standardized evaluation of extractive reading comprehension against curated Wikipedia passages, enabling systematic comparison of neural architectures.

Knowledge Base and Semantic Question Answering

Knowledge base QA (KBQA) answers questions by querying structured repositories such as Freebase, Wikidata, or corporate ontologies, translating natural language into formal graph queries over typed entities and relations. The challenge is semantic parsing: mapping an informal phrasing to a structured query expression whose precise logical form retrieves the correct answer. For linked data and RDF stores, this involves generating SPARQL queries that traverse subject-predicate-object triples. KBQA systems achieve high precision on well-formed factoid questions over covered entities but struggle with questions that require information absent from the knowledge base. The ACM Computing Surveys review on evaluating question answering systems examines how evaluation complexity increases when moving from factoid answering to complex multi-relation queries.

Neural and Retrieval-Augmented Approaches

Transformer-based language models, beginning with BERT in 2018, substantially advanced extractive QA performance by learning contextual token representations that capture long-range semantic dependencies unavailable to earlier models. More recently, retrieval-augmented generation (RAG) architectures combine a dense retrieval component with a large generative language model, allowing the system to draw on an external corpus at inference time while generating fluent, directly phrased answers rather than extracting spans verbatim. The arXiv survey on advances in natural language question answering traces the progression from rule-based systems through machine learning classifiers to end-to-end neural architectures and identifies the open challenges of multi-hop reasoning and domain adaptation.

Applications

Question answering has applications across a range of systems and domains, including:

• Virtual assistants in consumer devices and enterprise productivity tools
• Medical question answering for clinical decision support
• Scientific literature search for drug discovery and materials research
• Customer service automation and FAQ systems
• Legal and regulatory document querying for compliance analysis