[Paper Review] Text Generation Based on Generative Adversarial Nets with Latent Variable

Heng Wang, Zengchang Qin+1 more|arXiv (Cornell University)|Dec 1, 2017

Topic Modeling18 references3 citations

TL;DR

This paper proposes VGAN, a text generation model combining variational autoencoders (VAE) and generative adversarial networks (GAN) with a latent variable to improve diversity and realism. By integrating an RNN-based generator with VAE and a CNN discriminator, and training via policy gradient, VGAN achieves state-of-the-art performance on three benchmark datasets, outperforming RNNLM and SeqGAN in both negative log-likelihood and BLEU scores.

ABSTRACT

In this paper, we propose a model using generative adversarial net (GAN) to generate realistic text. Instead of using standard GAN, we combine variational autoencoder (VAE) with generative adversarial net. The use of high-level latent random variables is helpful to learn the data distribution and solve the problem that generative adversarial net always emits the similar data. We propose the VGAN model where the generative model is composed of recurrent neural network and VAE. The discriminative model is a convolutional neural network. We train the model via policy gradient. We apply the proposed model to the task of text generation and compare it to other recent neural network based models, such as recurrent neural network language model and SeqGAN. We evaluate the performance of the model by calculating negative log-likelihood and the BLEU score. We conduct experiments on three benchmark datasets, and results show that our model outperforms other previous models.

Motivation & Objective

• Address the limitations of standard GANs in text generation, particularly mode collapse and lack of diversity.
• Overcome the exposure bias and word-level training vs. sentence-level evaluation mismatch in RNN language models.
• Improve data distribution modeling in discrete sequence generation by introducing high-level latent random variables.
• Develop a unified framework combining VAE and GAN to enhance both diversity and realism in generated text.
• Demonstrate superior performance on standard text generation benchmarks using both likelihood and n-gram similarity metrics.

Proposed method

• Propose VGAN, a generative model where the generator combines an RNN with a VAE to model latent variables for better data distribution learning.
• Use a convolutional neural network (CNN) as the discriminator to distinguish real from generated text sequences.
• Train the generator using policy gradient to optimize the adversarial objective, enabling backpropagation through discrete outputs.
• Introduce a variational inference framework to approximate the true posterior over latent variables, enabling end-to-end training.
• Apply adversarial training with a reward signal derived from the discriminator to improve generator policy.
• Pre-train the generator using VAE objectives before adversarial fine-tuning to stabilize training and improve convergence.

Experimental results

Research questions

• RQ1 Can combining VAE with GAN improve the diversity and realism of generated text compared to standard GANs or RNNLMs?
• RQ2 Does the use of high-level latent variables in the generator help mitigate mode collapse and improve data distribution modeling in discrete sequence generation?
• RQ3 Can policy gradient training effectively optimize a discrete sequence generator when combined with adversarial training in a GAN framework?
• RQ4 How does the proposed VGAN model compare to strong baselines like RNNLM and SeqGAN in terms of negative log-likelihood and BLEU score across multiple datasets?
• RQ5 Does the pre-training stage of the generator (via VAE) lead to more stable and improved adversarial training in the VGAN framework?

Key findings

• VGAN achieves the lowest negative log-likelihood (NLL) of 191 on the Taobao dataset, outperforming RNNLM (219) and SeqGAN (212).
• On the Amazon dataset, VGAN achieves an NLL of 408, significantly lower than RNNLM (483) and SeqGAN (467).
• On the PTB dataset, VGAN achieves an NLL of 423, outperforming RNNLM (502), SeqGAN (490), and the pre-trained VGAN-pre (465).
• VGAN achieves the highest BLEU-2 score of 0.969 on the Taobao dataset, surpassing RNNLM (0.965), SeqGAN (0.968), and VGAN-pre (0.968).
• On the Amazon dataset, VGAN achieves a BLEU-2 score of 0.868, outperforming RNNLM (0.848), SeqGAN (0.856), and VGAN-pre (0.848).
• The adversarial training process leads to a consistent and significant drop in NLL across all datasets, indicating improved generator performance over time.