How2

How2数据集是由卡内基梅隆大学等机构创建的一个大规模多模态语言理解数据集，包含80,000个教学视频片段，总时长约2,000小时。该数据集不仅支持多模态分析，还提供了英语字幕及其众包葡萄牙语翻译，适用于多种语言处理任务。数据集的创建过程涉及从YouTube下载视频，提取视觉特征，并通过众包平台Figure Eight进行翻译。How2数据集的应用领域广泛，旨在解决机器在多模态环境下的语言理解和处理问题，促进语言、语音和视觉研究社区的协作。

The How2 dataset is a large-scale multimodal language understanding dataset created by Carnegie Mellon University and other institutions. It contains 80,000 instructional video clips, with a total duration of approximately 2,000 hours. This dataset not only supports multimodal analysis, but also provides English subtitles and their crowdsourced Portuguese translations, making it applicable to a variety of language processing tasks. The creation process of the dataset involves downloading videos from YouTube, extracting visual features, and carrying out translations via the crowdsourcing platform Figure Eight. The How2 dataset covers a wide range of application fields, aiming to solve the problems of machine language understanding and processing in multimodal environments, and promote collaboration among the language, speech and vision research communities.