Code-switched inspired losses for spoken dialog representations