Question Answering and Reading Comprehension

The evidence boundary and output contract determine what counts as success.

The boundary bites hardest in open-domain QA over questions people actually asked, and Natural Questions is built out of exactly those. Google released it in 2019. Its abstract states the provenance plainly: “Questions consist of real anonymized, aggregated queries issued to the Google search engine.”

The counts are the lesson. The training set holds 307,373 examples. Only 152,148 have a long answer and 110,724 have a short answer. Roughly half the questions are marked null, because the retrieved Wikipedia page does not answer them. That is not an adversarial split or a stress test bolted on afterwards. It is what a corpus of real queries looks like once somebody checks whether the evidence supports an answer at all. Every setting below inherits that arithmetic. They differ only in where evidence may come from and what the system is permitted to say back.

Each stage can fail while later output remains fluent.

Stage three is not a rare exception branch. In Natural Questions, roughly half of the 307,373 training examples end there: 152,148 have a long answer, 110,724 have a short answer, and the remainder are null.

A benchmark can reward pattern matching without the intended reading behavior.

Each item below is a way to score without doing the reading, and the size of such an effect is measurable. One distracting sentence, inserted into SQuAD paragraphs, took sixteen published models from an average of 75% F1 to 36%. Jia and Liang ran that test in 2017.

Extractive QA often uses exact match and token-level F1 against reference spans; valid alternative spans, normalization, multi-span answers, and semantically equivalent wording can complicate these measures.

Generative QA additionally needs factual support, completeness, attribution, contradiction, concision, and utility. Human or model-assisted review requires validated rubrics and a reviewer who checks the source.

The measure arrived with the dataset. SQuAD, published in 2016, holds “107,785 question-answer pairs on 536 articles”, and in it “the answer to each question is a segment of text from the corresponding reading passage”. A span is what makes exact match and token F1 computable at all. Its authors’ own baseline was “a strong logistic regression model, which achieves an F1 score of 51.0%”, against “human performance (86.8%)”. That gap is what the metric was built to expose. It says nothing about an answer that is correct in different words.

How much correctness those measures throw away has since been counted. In 2022 Bulian and colleagues collected 23,260 human answer-equivalence ratings over 17,655 candidate answers produced by BiDAF, XLNet and LUKE on SQuAD. Their paper opens with the diagnosis: “The predictions of question answering (QA) systems are typically evaluated against manually annotated finite sets of one or more answers. This leads to a coverage limitation that results in underestimating the true performance of systems, and is typically addressed by extending over exact match (EM) with predefined rules or with the token-level F1 measure.” The result gives the bias a direction. Exact match and token-level F1 systematically underestimate QA systems, because they reject answers a human judges equivalent to — or better than — the reference. The error runs one way, against the system. 23,260 human judgements is what it took to show it.

The set should expose the complete retrieval and answer decision.

Step three is where most evaluation sets are thin, and it does not require invention. Natural Questions reached its null cases by taking real queries and recording that the retrieved Wikipedia page did not answer them: 307,373 training examples, 152,148 with a long answer, 110,724 with a short answer. The unanswerable cases were counted, not simulated.

For ten answerable questions, create a minimally changed version that the corpus cannot support; preserve topic and wording so the difference depends on evidence, date, jurisdiction, or entity.

Measure answer rate, abstention precision, retrieval overlap, and claim support. Inspect whether the model answers from prior associations instead of the supplied corpus.

This exercise has been run in public, on a deployed system, by a reporter. New York City launched its MyCity chatbot in October 2023, on Microsoft's Azure AI services, to answer small-business questions. On 29 March 2024 Colin Lecher of The Markup published what came back when he asked it questions whose answers are fixed by the city's own law. Landlords, it said, need not accept Section 8 housing vouchers. On whether an employer may take a share of workers' tips: “Yes, you can take a cut of your worker's tips.” Both answers were fluent. Both were confidently wrong about the law the tool existed to explain. Neither came with any signal that the system had no authority to give them. The city left the tool online; Mayor Eric Adams acknowledged on 2 April 2024 that its answers were wrong in some areas.

Abstention is the other half of the same measurement, and its price can be read off a table. The legal benchmark reports: “As shown in Figure 4, Lexis+ AI provides accurate (i.e., correct and grounded) responses on 65% of queries, while Ask Practical Law AI refuses to answer queries 62% of the time and responds accurately just 18% of the time.” Thomson Reuters's Ask Practical Law AI declines most of what it is asked and is still accurate on only 18%. Silence is not by itself honesty. Answer rate and abstention precision have to be measured together, or a system can look careful and be wrong at the same time.