How2QA

要为视频质量检查任务收集How2QA，将同一组选定的视频剪辑显示给另一组AMT工作人员进行多选QA注释。每个工人都被分配一个视频片段，并被要求写一个问题，其中有四个答案候选者 (一个正确和三个干扰者)。同样，对工人隐藏叙述，以确保收集的QA对不会受到字幕的偏见。与TVQA类似，为每个问题的相关时刻提供了起点和终点。过滤低质量注释后，最终数据集包含44,007个QA对，从9035个视频中选择22k 60秒剪辑。

How2QA was collected for the video quality inspection task. The same set of pre-selected video clips were presented to another group of AMT workers for multiple-choice QA annotation. Each worker was assigned one video clip and required to compose a question along with four answer options: one correct answer and three distractors. To ensure the collected QA pairs were not biased by subtitle content, the transcripts were hidden from all workers. Similar to TVQA, start and end timestamps were provided for the relevant moment associated with each question. After filtering out low-quality annotations, the final dataset contains 44,007 QA pairs derived from 22,000 60-second clips selected from 9,035 source videos.