[论文解读] Unsupervised Text Style Transfer using Language Models as Discriminators

Zichao Yang, Zhiting Hu+3 more|arXiv (Cornell University)|May 30, 2018

Topic Modeling参考文献 44被引用 47

一句话总结

本文通过用目标领域语言模型替代二元判别器来引导无监督文本风格转换，使端到端训练实现连续近似，并在解码、情感修改及相关翻译任务上取得改进。

ABSTRACT

Binary classifiers are often employed as discriminators in GAN-based unsupervised style transfer systems to ensure that transferred sentences are similar to sentences in the target domain. One difficulty with this approach is that the error signal provided by the discriminator can be unstable and is sometimes insufficient to train the generator to produce fluent language. In this paper, we propose a new technique that uses a target domain language model as the discriminator, providing richer and more stable token-level feedback during the learning process. We train the generator to minimize the negative log likelihood (NLL) of generated sentences, evaluated by the language model. By using a continuous approximation of discrete sampling under the generator, our model can be trained using back-propagation in an end- to-end fashion. Moreover, our empirical results show that when using a language model as a structured discriminator, it is possible to forgo adversarial steps during training, making the process more stable. We compare our model with previous work using convolutional neural networks (CNNs) as discriminators and show that our approach leads to improved performance on three tasks: word substitution decipherment, sentiment modification, and related language translation.

研究动机与目标

• Motivate unsupervised text style transfer without parallel data.
• Disentangle content and style to enable style transfer while preserving content.
• Introduce a language-model-based discriminator to provide stable token-level feedback.
• Develop a continuous approximation for back-propagation through discrete text generation.
• Evaluate on decipherment, sentiment alteration, and related language translation tasks.

提出的方法

• Model uses an encoder to extract content and a style-conditioned decoder to generate transferred text.
• Replace traditional binary discriminators with a target-domain language model that scores sentence likelihood.
• Define LM-based loss L_LM with real and transferred sentences to guide generation (Eq. 1–3).
• Apply a continuous Gumbel-softmax approximation to enable back-propagation through discrete tokens (Eq. 5).
• Two-step training: update language models with real and transferred data, then update the encoder/decoder with reconstruction and LM-based transfer loss.
• Explore the role of negative samples and show that omitting adversarial steps improves stability.

实验结果

研究问题

• RQ1 Can a language model trained on target-domain data provide a more stable and informative signal than a binary classifier for style transfer?
• RQ2 Does end-to-end training with a continuous token-level feedback mechanism improve fluency and style accuracy across tasks?
• RQ3 How does LM-based discrimination compare to CNN-based discriminators in decipherment, sentiment modification, and related translation?
• RQ4 What is the impact of using negative samples in LM-based training on stability and performance?

主要发现

• The LM-based discriminator achieved strong decipherment BLEU scores, e.g., 89.0% at 20% change and 49.3% at 100% change (LM row in Table 1).
• LM-based models outperform CNN-discriminator baselines in sentiment modification on accuracy and BLEU with better fluency (e.g., LM: 83.3 Accu, 38.6 BLEU; LM + Classifier: 91.2 Accu, 57.8 BLEU in Table 2).
• In sentiment tests, LM-based approaches yielded lower perplexities than several baselines (e.g., PPL_X 30.3 and PPL_Y 42.1 for LM; Table 2).
• Related-language translation results show LM-based methods achieving BLEU scores of 81.6 (sr–bs) and 85.5 (tw–cn), outperforming Shen et al. (2017) (Table 4).
• Using the LM alone (without adversarial steps) provides more stable training and competitive performance across tasks.
• Combining LM with a classifier can further improve sentiment transfer by balancing style modification and fluency (LM + Classifier in Table 2).