On the ability of monolingual models to learn language-agnostic representations