One reason is that the amount of training data is many many orders of magnitude smaller. FWIW it seems the structure you're talking about exploiting is at a mo

One reason is that the amount of training data is many many orders of magnitude smaller. FWIW it seems the structure you're talking about exploiting is at a morphological and syntactic level, which modern language models tend to effectively handle. Semantics are a much harder problem.