We used the YouTube Data API to augment the YouTube 8M corpus by crawling a variety of meta data for the videos.

First point of interest was the "video resource," which comprises data about the video, such as the video's title, description, uploader name, tags, view count, and more. Also included in the meta data is whether comments have been left for the video. If so, we downloaded them as well, including information about their authors, likes, dislikes, and responses.

There is no property which specifies a video's language, since this information is not mandatory when uploading a video. Also, the API provides only information about the available captions, but not the captions themselves. Only the uploader of a video is given access to its captions via the API; we extracted them using youtube-dl. For each video, all manually created captions were downloaded, and auto-generated captions in the "default" language and English. The "default" auto-generated caption gives perhaps the only hint at a video's original language.

Finally, we downloaded all thumbnails used to advertise a video, which are not available via the API, but only via a canonical URL. Our corpus provides the possibility to recreate the way a video is presented on YouTube (meta data and thumbnail), what the actual content is ((sub)titles and descriptions), and how its viewers reacted (comments).


Please refer to this publication for citing the dataset.

  • Download the dataset from Zenodo.
  • Find the related metadata at Google.