Selective Differential Privacy for Language Modeling