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