Escaping the sentence-level paradigm in machine translation