Introduction
Language style transfer—rewriting text to match a specific persona, formality level, or author's voice while preserving original meaning—typically requires massive amounts of parallel (labeled) data. In real-world scenarios, we often need to transfer a single source text into multiple different styles, but parallel data across all these styles is rarely available.
Previous approaches often treat each style pair as an isolated task or attempt standard multi-task learning by sharing model parameters. However, parameter sharing can actually hurt performance in style transfer because distinct target styles may conflict, leading to "target inconsistency." This paper introduces a Multi-Agent Style Transfer (MAST) system that leverages mutual similarities between styles to train highly accurate systems with minimal labeled data.
Methodology
The authors propose a framework that moves away from multi-task parameter sharing toward a collaborative multi-agent approach.
- Semi-Supervised Base Models: Individual one-to-one style transfer models are built using an Attentional Sequence-to-Sequence (AttS2S) architecture. These are enhanced with unlabeled data via Denoising Auto-Encoders (DAE) and Back-translation.
- Multi-Agent Style Transfer (MAST): Each pre-trained model is treated as an "agent." During text generation, these agents collaborate to predict the next word.
- Self-Organization by Similarities (SOS): Agents select "neighbor" agents based on target style similarity (measured via a binary classifier) and individual performance (BLEU scores).
- Multi-Agent Training (MAT): A lightweight RNN
Controlleris trained for each agent to dynamically predict a weight distribution over the agent and its neighbors, combining their vocabulary probability distributions.
Experimental Evaluation
The system was tested using 15 different English translations of the Bible, with the American Standard Version (ASV) as the source. To simulate data scarcity, labeled data was restricted to just 2,000 aligned sentence pairs per style, alongside 20,000 unaligned sentences for unlabeled training.
Key Results and Insights
The MAST system demonstrated significant improvements over all baselines. On a 5-target task, the MAST:SOS-2 configuration achieved an average BLEU score of 29.84, outperforming the vanilla AttS2S baseline (23.27) and the standard multi-task learning approach (MulT, 24.96).
Key Takeaways:
- Collaboration Over Sharing: Treating models as independent agents that "vote" on the next word effectively bypasses the multi-task learning bottleneck, allowing models to remain specialized while benefiting from peer knowledge.
- Selective Listening: The success of the system relies heavily on the SOS algorithm. Collaborating with randomly selected neighbors (
MAST:Rand-k) proved unstable, highlighting the importance of selecting neighbors with similar styles and high accuracy. - Scalability and Robustness: Performance continued to improve as the pool of target styles grew from 5 to 15, as agents could find better matches. Furthermore, the system remained robust even when labeled data was reduced to only 500 parallel sentences.
Conclusion and Future Work
MAST provides a scalable and effective framework for multi-style transfer under low-resource conditions. While highly effective, the system is sensitive to neighbor selection. Future work will explore more advanced algorithms for identifying helpful neighbors and investigate alternative architectures for the Controller network to further push the performance ceiling.
Summary updated