Up to five questions that were impossible to answer based on each paragraph alone were written with the help of crowd workers, referring entities in the paragraph and ensuring that a credible response could be found from the entire SQuAD 1.1 text. Those questions from workers who didn't understand the task on that article were screened to reduce noise. The train, development, and test splits were constructed using the same article division as SQuAD 1.1, with each split integrating old and new data and eliminating articles containing unanswerable questions. For SQuAD 2.0, the ratio of answerable to unanswerable questions in these splits is around one-to-one, whereas train data has roughly twice as many answerable as unanswerable questions.