I'm not sure how this would translate into a query, but have you thought about analyzing the longest chain of devices and comparing their PDFs? I feel like that would potentially demonstrate the game-of-telephone effect you're getting at.
Fig 2 [2] shows a graph of similarity of all devices to their predicate for each year.
However the discussion sections mentions difficulty with using Word2Vec because of the jargon used to describe these devices, so I'm not totally sure how to interpret the graph.
[1]: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8496833/
[2]: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8496833/figure/...