Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage, or the question might be unanswerable.
SQuAD2.0 combines the 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions written adversarially by crowdworkers to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but also determine when no answer is supported by the paragraph and abstain from answering.

Data Collection

We employed crowd workers on the Daemo crowd-sourcing platform to write unanswerable questions. Each task consisted of an entire article from SQuAD 1.1. For each paragraph in the article, workers were asked to pose up to five questions that were impossible to answer based on the paragraph alone, while referencing entities in the paragraph and ensuring that a plausible answer is present. As inspiration, we also showed questions from SQuAD 1.1 for each paragraph; this further encouraged unanswerable questions to look similar to answerable ones.
We removed questions from workers who wrote 25 or fewer questions on that article; this filter helped remove noise from workers who had trouble understanding the task, and therefore quit before completing the whole article. We applied this filter to both our new data and the existing answerable questions from SQuAD 1.1. To generate train, development, and test splits, we used the same partition of articles as SQuAD 1.1, and combined the existing data with our new data for each split. For the SQuAD 2.0 development and test sets, we removed articles for which we did not collect unanswerable questions. This resulted in a roughly one-to-one ratio of answerable to unanswerable questions in these splits, whereas the train data has roughly twice as many answerable questions as unanswerable ones. To confirm that our dataset is clean, we hired additional crowd workers to answer all questions in the SQuAD 2.0 development and test sets. In each task, we showed workers an entire article from the dataset. For each paragraph, we showed all associated questions; unanswerable and answerable questions were shuffled together. For each question, workers were told to either highlight the answer in the paragraph, or mark it as unanswerable. Workers were told to expect every paragraph to have some answerable and some unanswerable questions. They were asked to spend one minuteper question, and were paid $10.50 per hour.
To reduce crowd worker noise, we collected multiple human answers for each question and selected the final answer by majority vote, breaking ties in favor of answering questions and preferring shorter answers to longer ones. On average, we collected 4.8 answers per question.