Maybe you could use a language model embedding to define some kind of semantic distance.