The limitations of architectures that solely rely on correlational information, like those without any built-in knowledge or pre-existing concepts, in solving inference and understanding problems in natural language processing.