Comparing the representation learning of autoencoding transformer models in ad hoc information retrieval