adding knowledge to language models