Methodology
The information analyzed for this text consists of movies, metadata and social media posts collected from Twitter, YouTube and Pomegranate Cloud between Feb. 18 and June 2. We downloaded greater than 5,000 movies posted to those platforms between Jan. 23 (the date of the primary marketing campaign video following the State Division’s Jan. 19 declaration of genocide in Xinjiang) and Could 31. On Pomegranate Cloud, we collected clips focusing on Mr. Pompeo by trying to find posts mentioning him after Jan. 19 that contained video. We collected movies denying pressured labor in Xinjiang’s cotton trade from a piece devoted to them within the app.
On Twitter and YouTube, the marketing campaign movies had been collected from what we name “warehouse accounts,” these whose movies had been shared by a community of greater than 300 coordinated Twitter accounts. This community appeared to work in coordination to love and retweet content material that supported Chinese language authorities insurance policies, such because the marketing campaign movies, in addition to information articles and editorials from state media. To outline the community, we manually recognized a small group of accounts and their indicators of automation, particularly posts containing similar content material adopted by strings of random characters. We then recognized extra community accounts by programmatically trying to find different accounts that boosted the identical content material and had the identical indicators.
We cataloged greater than 3,000 distinctive marketing campaign movies out of the greater than 5,000 collected. To pinpoint duplicates amongst movies containing varied compression charges, visible artifacts and subtitle languages, we calculated a fingerprint for every video by working a pattern of its frames by way of the Google Cloud Imaginative and prescient picture labeler. We decided movies with comparable fingerprints and durations to be duplicates. We manually sampled and reviewed the outcomes from this course of to attenuate false positives and false negatives.
To establish non-campaign movies on YouTube and Twitter, we first obtained their subtitles through the use of optical character recognition on frames taken from them at common intervals. We thought of movies that didn’t point out Mr. Pompeo or cotton to be non-campaign movies. Movies on Twitter and YouTube at all times had subtitles in English and Chinese language. Movies from Pomegranate Cloud had subtitles in Chinese language solely when Uyghur was spoken. We thought of all movies collected from Pomegranate Cloud to be marketing campaign movies, however didn’t assessment every manually. We additionally used the subtitles of the YouTube and Twitter movies to investigate their content material.