Systematic generalisation, in a nutshell, works like this:
one apple, two apples
one ball, two balls
one rose, two roses
one ___, two ___s
It’s an actual feature of language, and it operates on both the morphological and syntactical layers.
And IMO a good start, but not enough. As machine text generation moves away from LLMs and their “ooga booga, bash token on token” approach, eventually you’ll need to deal with the fact that the morpheme (aka token) itself don’t matter that much, it’s just an interface for a semantic layer. And that you need that semantic layer if you want anything past “potatoes are active, oranges are passive”.
Systematic generalisation, in a nutshell, works like this:
It’s an actual feature of language, and it operates on both the morphological and syntactical layers.
And IMO a good start, but not enough. As machine text generation moves away from LLMs and their “ooga booga, bash token on token” approach, eventually you’ll need to deal with the fact that the morpheme (aka token) itself don’t matter that much, it’s just an interface for a semantic layer. And that you need that semantic layer if you want anything past “potatoes are active, oranges are passive”.