“Average” Approximates “First Principal Component”? An Empirical Analysis on Representations from Neural Language Models