Text Style Transfer (TST) aims to rewrite text in a target style while preserving the original meaning. CycleGAN architectures have shown state-of-the-art performance on benchmark TST data, but potentially suffer from the lack of domain specific supervision. In this paper, we study the robustness of CycleGANs to domain adaptation for TST. Specifically, we first adapt a sequence-level CycleGAN architecture to two particularly challenging domains, such as Shakespearean–modern English rewriting and toxic-to-neutral detoxification, and then study how the architecture components influence the performance of an unsupervised domain adaptation pipeline. Results show that domain-adapted CycleGANs achieve strong content preservation and competitive balanced performance against existing unsupervised methods. Notably, CycleGANs consistently outperform Large Language Models (LLMs) in content preservation under zero-shot settings, and continue to outperform most LLMs even under few-shot prompting, while remaining competitive in terms of style accuracy.

Domain-Adaptive CycleGAN Training for Text Style Transfer / Ibrahim, H., Gallipoli, G., Cagliero, L.. - (In corso di stampa). (20th Anniversary IEEE International Conference on Application of Information and Communication Technologies (AICT) Baku (AZE) October 14–16, 2026).

Domain-Adaptive CycleGAN Training for Text Style Transfer

Ibrahim, Hadi;Gallipoli, Giuseppe;Cagliero, Luca
In corso di stampa

Abstract

Text Style Transfer (TST) aims to rewrite text in a target style while preserving the original meaning. CycleGAN architectures have shown state-of-the-art performance on benchmark TST data, but potentially suffer from the lack of domain specific supervision. In this paper, we study the robustness of CycleGANs to domain adaptation for TST. Specifically, we first adapt a sequence-level CycleGAN architecture to two particularly challenging domains, such as Shakespearean–modern English rewriting and toxic-to-neutral detoxification, and then study how the architecture components influence the performance of an unsupervised domain adaptation pipeline. Results show that domain-adapted CycleGANs achieve strong content preservation and competitive balanced performance against existing unsupervised methods. Notably, CycleGANs consistently outperform Large Language Models (LLMs) in content preservation under zero-shot settings, and continue to outperform most LLMs even under few-shot prompting, while remaining competitive in terms of style accuracy.
In corso di stampa
File in questo prodotto:
File Dimensione Formato  
cyclegan.pdf

accesso riservato

Descrizione: Camera-Ready / Author's Accepted Manuscript
Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: Non Pubblico - Accesso privato/ristretto
Dimensione 295.61 kB
Formato Adobe PDF
295.61 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3017951