“…Other unsupervised methods for paraphrase generation include VAE (VQ-VAE) (Roy and Grangier, 2019), latent bagof-words alignment (Fu et al, 2019) and simulated annealing (Liu et al, 2019a). Adapting large-scale pretraining (Devlin et al, 2018;Radford et al, 2018;Liu et al, 2019b;Clark et al, 2020;Sun et al, 2021b) to paraphrase generation has been recently investigated (Witteveen and Andrews, 2019;Hegde and Patil, 2020;Niu et al, 2020;Meng et al, 2021) and has shown promising potentials to improve generation quality. Our work is distantly related to unsupervised text style transfer (Hu et al, 2017;Mueller et al, 2017;Shen et al, 2017;Li et al, 2018a;Fu et al, 2018), where the model alters a specific text attribute of an input sentence (such as sentiment) while preserving other text attributes.…”