What you say is correct but it doesn’t explain why the model fails to capture the semantics of “is”.