NLP Basics Q&A

Answer: QA is the task of taking a natural language question and returning an appropriate answer, often by reading a context passage (reading comprehension) or using a knowledge base (open-domain QA).

Answer: Extractive QA selects a span of text directly from the context as the answer, whereas abstractive QA may generate a new answer sentence that paraphrases or summarizes the relevant information.

Answer: SQuAD (Stanford Question Answering Dataset) is a large-scale reading comprehension benchmark where models read Wikipedia passages and answer crowd-sourced questions, and it became a standard for evaluating extractive QA models.

Answer: A span-based QA model predicts the start and end positions of the answer within the context passage, effectively treating QA as span extraction rather than free-form generation.

Answer: For extractive QA, BERT encodes the concatenated question and context, and two classifier heads predict start and end indices over the context tokens to mark the answer span.

Answer: Open-domain QA answers questions using a very large corpus or the web instead of a single given passage, usually combining a retriever (to find documents) and a reader (to extract or generate answers).

Answer: In closed-book QA the model must answer using only its internal parameters (no external context at inference time), while in open-book QA it is given supporting documents or can retrieve them at inference time.

Answer: QA models often predict start and end positions at the token level, so consistent tokenization (e.g. WordPiece or BPE) is critical to align model outputs with the original text spans.

Answer: In some datasets the correct behavior is to say that the context does not contain an answer; models handle this by predicting a special no-answer score or a span that maps to “no answer”.

Answer: Exact Match (EM) and token-level F1 are standard metrics, measuring whether the predicted span matches the gold span exactly and how much the token overlap is between them.

Answer: Attention allows the model to focus on question-relevant parts of the context, aligning question tokens with context tokens to better locate the answer span or generate an answer.

Answer: Multi-hop QA requires reasoning over multiple passages or sentences to answer a question, whereas single-hop QA can be solved using one local piece of evidence.

Answer: A retriever first fetches candidate documents from a large corpus (e.g. using BM25 or dense retrieval), and then a reader model performs fine-grained extraction or generation over the retrieved passages.

Answer: Normalization (e.g. lowercasing, removing punctuation and articles) reduces spurious mismatches so that semantically identical answers like “the Eiffel Tower” and “Eiffel Tower” are treated as equivalent.

Answer: Large language models can be prompted with the question and optionally context to directly generate a natural language answer, often combining retrieval with generation for factual accuracy.

Answer: Typical errors include misunderstanding the question, attending to the wrong part of the context, predicting partially correct spans, and hallucinating unsupported answers in generative settings.

Answer: Factoid QA expects short, factual answers (names, dates, entities), while non-factoid QA may require longer, explanatory answers such as “why” or “how” questions.

Answer: QA models trained on one domain (like Wikipedia) can degrade on another (like biomedical text), so adapting or fine-tuning on in-domain data improves robustness and accuracy.

Answer: Conversational QA handles multi-turn dialogues where each question may depend on previous turns, requiring the model to track context, coreference and dialogue history.

Answer: RAG-style models combine a neural retriever with a generative model so answers are grounded in retrieved documents, improving factual correctness while keeping generation flexible.