It lets the decoder use both preceding and following context to generate more fluent, natural summaries.