Text this: Scalable syntactic inductive biases for neural language models