The linked tweet has a diagram where you can pretty quickly see that this isn't just about using wikidata as a training set. The paper linked from the tweet also gives a good summary on its first page.