Keeping an Eye on Context: Attention Allocation over Input Partitions in Referring Expression Generation