RLSeq2Seq

Why we included this project

Researchers and graduate students working on text generation will find this repository useful: it is the reference implementation behind the IEEE TNNLS paper on deep reinforcement learning for sequence-to-sequence models, so it is a good starting point for reproducing or extending those experiments. The code packs several training strategies into one TensorFlow codebase, including scheduled sampling, end-to-end backprop, policy gradient with self-critic, and actor-critic variants built on DDQN and dueling networks, which makes it a convenient way to see how different RL objectives behave on the same abstractive summarization task. Helper scripts download and preprocess the CNN/Daily Mail and Newsroom datasets, so you can compare ROUGE scores against the paper without much setup. One caveat: it targets Python 2.7 and TensorFlow 1.10 and is no longer actively maintained, so treat it as a study resource and a research baseline rather than something to run in production.