> All of these will perform differently on the negation issue. That's a bit of a cop-out, but the classical and logical (in the mathematical sense) way to do i

> All of these will perform differently on the negation issue. That's a bit of a cop-out, but the classical and logical (in the mathematical sense) way to do it, is to have a feature that represents negation. Embeddings don't work like this, so it's quite possible that the presence of negation gets pushed aside in the processing of the answer, simply because other features matter more. It's also to be expected that such models will have problems with similar abstract concepts that have more effect on the interpretation than their "physical" presence would suggest, such as nested existential qualifiers, and consequently logical proof. By enlarging the model, you can fake it a bit.