[Paper Review] Style Transfer from Non-Parallel Text by Cross-Alignment

Tianxiao Shen, Tao Leí+2 more|arXiv (Cornell University)|May 26, 2017

Topic Modeling25 references353 citations

TL;DR

The paper presents a cross-alignment framework to perform style transfer using non-parallel text by learning a shared content latent space and aligning distributions across styles, evaluated on sentiment modification, decipherment of word substitution ciphers, and word-order recovery.

ABSTRACT

This paper focuses on style transfer on the basis of non-parallel text. This is an instance of a broad family of problems including machine translation, decipherment, and sentiment modification. The key challenge is to separate the content from other aspects such as style. We assume a shared latent content distribution across different text corpora, and propose a method that leverages refined alignment of latent representations to perform style transfer. The transferred sentences from one style should match example sentences from the other style as a population. We demonstrate the effectiveness of this cross-alignment method on three tasks: sentiment modification, decipherment of word substitution ciphers, and recovery of word order.

Motivation & Objective

• Motivate style transfer in settings where only non-parallel, mono-style data are available.
• Propose a refined cross-alignment approach that disentangles content from style without heavily perturbing content representations.
• Demonstrate effectiveness across multiple tasks—sentiment modification, word substitution decipherment, and word order recovery—using non-parallel data.
• Explore theoretical and practical conditions under which content distributions can be aligned across styles to enable transferable mappings.

Proposed method

• Introduce an encoder E that maps a sentence and its style to a style-independent content representation z.
• Introduce a style-dependent generator G that decodes x from a given style y and content z, enabling cross-style transfer.
• Use two training variants: aligned auto-encoder (with adversarial discriminator to align z distributions across styles) and cross-aligned auto-encoder (aligns transferred samples with real samples from the other style using dual discriminators).
• Avoid enforcing a simple p(z) prior; instead, encourage rich content representations and cross-population distributional alignment to support non-parallel style transfer.
• Apply continuous relaxation and Professor-Forcing to train with discrete language outputs, mitigating gradient estimation issues.
• Evaluate with two discriminators and two training signals to ensure latent space alignment and robust cross-domain transfer.

Experimental results

Research questions

• RQ1 Can non-parallel corpora with the same content distribution be used to learn style transfer mappings by aligning latent content across styles?
• RQ2 Does cross-alignment of latent representations improve content preservation and style transfer quality across sentiment modification, decipherment, and word order tasks?
• RQ3 What are the practical training strategies to handle discrete language generation and align distributions without explicit p(z) priors?

Key findings

• Cross-aligned auto-encoder achieves higher sentiment transfer quality than a non-cross-aligned baseline, with a reported sentiment transfer accuracy of 78.4% on transferred sentences.
• Aligned auto-encoder and cross-aligned auto-encoder show substantial improvements in non-parallel decipherment tasks, achieving higher BLEU scores than aligned or VAE baselines across substitution and word-order challenges.
• In sentiment evaluation, cross-aligned auto-encoder attains an overall transfer score of 41.5% with a fluency score of 2.8 on human evaluation, compared to 41.0% for the baseline control-gen model in some settings.
• On sentiment classification, a pretrained classifier reports 83.5% accuracy for Hu et al. (2017)’s method, while cross-aligned auto-encoder achieves 78.4% accuracy in the authors’ setup.
• In decipherment experiments, cross-aligned auto-encoder consistently outperforms non-cross-aligned baselines across multiple substitution rates (e.g., 20%–100%), with BLEU scores reaching 57.4 for 100% substitution and 26.1 for word order recovery when compared to non-transfer baselines.
• The results demonstrate that cross-alignment strengthens latent z alignment and improves transfer quality across diverse non-parallel style transfer tasks.