Presumably, they'd want to get at embeddings, and compare the dimensional space somehow to say: 'the relation between tokens a,b,c is close to the relation of t

Presumably, they'd want to get at embeddings, and compare the dimensional space somehow to say: 'the relation between tokens a,b,c is close to the relation of tokens a1,b1,c1 in a similar model of texts of known language of apparently same family (same up to aN,bN,cN), and out of these N sequences, sequence X makes most sense given existing examples'. (As you can tell, the argument involves some handwaving, but it may possible?)