How2QA is a video + language learning framework dataset. The dataset presents the same set of selected video clips to another set of AMT workers for multiple-choice question-answer annotation. Each worker is assigned a video clip and asked to write a question based on four prepared responses (one correct answer and three distracting answers). The video narration is hidden from the workers to ensure that the collected question-answer pairs are not affected by subtitles. The dataset contains 22,000 60-second clips and 44,007 question-answer pairs selected from 9,035 videos.