Understanding Long Documents with Different Position-Aware Attentions