| Summary: | 碩士 === 國立成功大學 === 資訊工程學系 === 106 === With the significant development of the deep learning technique, more and more neural networks have been proposed to solve some intricate problems about sequence generation. Generative adversarial Learning is one of the most novel strategies. Models applying this particular idea generally called “Generative Adversarial Nets (GANs)” consists of two parts: the Generator and the Discriminator. These models use the discriminator to guide the training of the generator for improving the effectiveness of each other. At the same time, GANs have already achieved great contributions to image processing. However, the effect of GANs on text generation has been shown unstable. Three major limitations cause GANs are hard to make a breakthrough in Nature Language Processing (NLP). Firstly, with considering the dialogue generation problem as a kind of decision-making step, the discrete outputs generated by the sampling operation is difficult to pass through the gradient from the discriminator to the generator. Secondly, prediction errors will be accumulated during generating sequence because of the different strategy between training and testing using the recurrent neural network (RNN). Therefore, we call it “exposure bias” for short. Finally yet importantly, the discriminator is only able to evaluate a complete sequence, which for every time steps, it is harsh that to extract the current score for every partial word. In summary, how to deal with these series of questions has become the critical factor if we can apply GANs in the NLP field.
In this paper, we propose a conditional sequence generative adversarial network to solve these problems by using the attention-based reward strategy. We jointly train an attention mechanism and the GANs. This model dynamically assigns the weights of feedback information from the discriminator back to the generator conditioned on the potential associations between words and sentences, which makes the training process much more stable and computationally efficient. Experimental results on synthetic data demonstrate that our model can generate better sequences. Moreover, we report a significant improvement of our model over the previous baselines on several real-world tasks.
|
|---|
