Overestimation of Syntactic Representation in Neural Language Models