Oh yeah, I do mean Norwegian there. I have made changes to the article.
It is “supervised”, in fact (‘self-supervised’ to be precise). The Skip-Thought encoder, by design, only requires raw text in the input. The supervision arises from the correct ordering of sentences that we provide in the input. It may seem confusing because we don’t give any specific ‘labels’ in the input, but it is, in fact, ‘supervised’. It’s just that this kind of ‘supervised’ data is available in abundance on the Internet.