A PyTorch implementation of the model from "An End-to-End Generative Architecture for Paraphrase Generation", (Qian Yang et al), from EMNLP 2019, http://people.ee.duke.edu/~lcarin/emnlp_gap.pdf (a try). The policy gradients implemented use a single reward for the entire sentence. You are encouraged to raise any doubts regarding the working of the code as an issue.