Anfonwch hwn fel neges destun: Scalable syntactic inductive biases for neural language models