[Paper Review] Style Transfer as Unsupervised Machine Translation

Zhirui Zhang, Shuo Ren+6 more|arXiv (Cornell University)|Aug 23, 2018

Natural Language Processing TechniquesComputer Science30 references114 citations

TL;DR

The paper reframes style transfer as an unsupervised machine translation problem, building word-level style mappings, SMT-based transfer systems, and iteratively trained NMT models with a style classifier to improve style accuracy and content preservation without parallel data.

ABSTRACT

Language style transferring rephrases text with specific stylistic attributes while preserving the original attribute-independent content. One main challenge in learning a style transfer system is a lack of parallel data where the source sentence is in one style and the target sentence in another style. With this constraint, in this paper, we adapt unsupervised machine translation methods for the task of automatic style transfer. We first take advantage of style-preference information and word embedding similarity to produce pseudo-parallel data with a statistical machine translation (SMT) framework. Then the iterative back-translation approach is employed to jointly train two neural machine translation (NMT) based transfer systems. To control the noise generated during joint training, a style classifier is introduced to guarantee the accuracy of style transfer and penalize bad candidates in the generated pseudo data. Experiments on benchmark datasets show that our proposed method outperforms previous state-of-the-art models in terms of both accuracy of style transfer and quality of input-output correspondence.

Motivation & Objective

• Motivate style transfer as a translation task to leverage unsupervised learning without parallel data.
• Develop a two-stage training pipeline: (1) word-level transfer table construction and SMT-based systems; (2) bidirectional NMT with iterative back-translation.
• Incorporate a style classifier to control noise and ensure transfer accuracy during training.
• Demonstrate improved style transfer accuracy and content preservation over prior state-of-the-art methods on benchmark datasets.

Proposed method

• Construct a word-level style transfer table P_{s→t}(y_w|x_w) using word embedding similarity and style preference signals.
• Build bidirectional SMT transfer systems using the transfer tables and style-specific language models to generate pseudo-parallel data.
• Initialize bidirectional NMT-based transfer models from the pseudo data and train them with iterative back-translation.
• Introduce an external style classifier to reward high-style-accuracy outputs and penalize bad pseudo data during back-translation optimization.
• Formalize the training objective as a combination of a forward-backward likelihood lower bound and a classifier-based reward term.
• Use beam search to approximate expectations and sample top-k translations for training (Algorithm 1).

Experimental results

Research questions

• RQ1 Can non-parallel, same-language corpora be leveraged to learn effective style transfer without parallel data?
• RQ2 Does a two-stage approach (word-level transfer plus iterative back-translation) improve over purely auto-encoder or retrieve-based methods?
• RQ3 What is the impact of adding a style classifier on the stability and quality of back-translation based training?
• RQ4 How does the proposed method perform across different style transfer tasks (sentiment in Yelp/Amazon, romantic/humorous captions) compared to baselines?

Key findings

• The proposed method outperforms state-of-the-art baselines on style transfer accuracy and content preservation across Yelp, Amazon, and Captions datasets.
• Word-level transfer tables combined with style-specific language models enable effective initial SMT-based transfer, which informs subsequent NMT-based training.
• Iterative back-translation with a style classifier yields mutual improvements between bidirectional models and stabilizes training.
• Human and automatic evaluations show the proposed approach achieves higher style transfer accuracy and better input-output quality than CrossAligned, MultiDecoder, StyleEmbedding, TemplateBased, and Del-Retr-Gen.
• A style classifier is crucial; removing it degrades transfer accuracy and overall performance.
• Analysis indicates the SMT stage provides strong initial pseudo data, while iterative NMT with back-translation optimizes content fidelity and fluency.