How do neural machine translation models differ from traditional rule-based or statistical translation methods in terms of accuracy and efficiency?

How do neural machine translation models differ from traditional rule-based or statistical translation methods in terms of accuracy and efficiency?