BoFGAN: Towards A New Structure of Backward-or-Forward Generative Adversarial Nets

Natural Language Generation (NLG), as an important part of Natural Language Processing (NLP), has begun to take full advantage of recent advances in language models. Based on recurrent neural networks (RNNs), NLG has made ground breaking improvement and is widely applied in many tasks. RNNs typically learn a joint probability of words, and the additional information is usually fed to RNNs hidden layer using implicit vector representations. Still, there exists some problem unsolved. Standard RNN is not applicable when we need to impose hard constraints on the language generation tasks: for example, standard RNNs cannot guarantee designated word(s) to appear in a target sentence to generate. In this paper, we propose a Backward-or-Forward Generative Adversarial Nets model (BoFGAN) to address this problem. Starting from a particular given word, a generative model at every time step generates a new preceding or subsequent word conditioned on the generated sequence so far until both sides reach an end. To train the generator, we first model it as a stochastic policy using Reinforcement Learning; then we employ a discriminator to evaluate the quality of a complete sequence as the end reward; and lastly, we apply Monte Carlo (MC) search to estimate the long-term return and update the generator via policy gradient. Experimental results demonstrate the effectiveness and rationality of our proposed BoFGAN model.