Rethinking Style Transformer with Energy-based Interpretation: Adversarial Unsupervised Style Transfer using a Pretrained Model

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5452 - 5467 December 7-11, 2022 ©2022 Association for Computational Linguistics Rethinking Style Transformer with Energy-based Interpretation: Adversarial Unsupervised Style Transfer using a Pretrained Model Hojun Cho1, Dohee Kim1, Seungwoo Ryu1, ChaeHun Park1, Hyungjong Noh2, Jeong-in Hwang2, Minseok Choi1, Edward Choi1, and Jaegul Choo1 1KAIST AI 2NCSOFT Corporation {hojun.cho,dohee1121,swryu,ddehun,minseok.choi,edwardchoi,jchoo}@kaist.ac.kr {nohhj0209,jihwang}@ncsoft.com Abstract Style control, content preservation, and fluency determine the quality of text style transfer mod- els. To train on a nonparallel corpus, several existing approaches aim to deceive the style dis- criminator with an adversarial loss. However, adversarial training significantly degrades flu- ency compared to the other two metrics. In this work, we explain this phenomenon us- ing energy-based interpretation, and leverage a pretrained language model to improve flu- ency. Specifically, we propose a novel ap- proach which applies the pretrained language model to the text style transfer framework by restructuring the discriminator and the model itself, allowing the generator and the discrim- inator to also take advantage of the power of the pretrained model. We evaluated our model on three public benchmarks GYAFC, Amazon, and Yelp and achieved state-of-the-art perfor- mance on the overall metrics. 1 Introduction Text style transfer is the task of converting a sen- tence from one style to another while preserving style-agnostic semantics. In solving the text style transfer task, three criteria must be considered: 1) style control, how well a style is transferred from the original sentence to the generated one, 2) con- tent preservation, how well the generated sentence has retained the semantics of the original, and 3) fluency, how natural the generated sentence is. Text style transfer is challenging, since fluently converting the style of a sentence often conflicts with content preservation (Prabhumoye et al., 2018; John et al., 2019; Gong et al., 2019). To address this challenge, several supervised text style trans- fer methods have been attempted (Jhamtani et al., 2017; Al Nahas et al., 2019; Lai et al., 2021b). But style-labeled sentence pairs are often not avail- able, making that approach less practical in a real- world setting. Various unsupervised text style Style 𝑠𝑠′ Style 𝑠𝑠 Corpus 𝒙𝒙 ∇𝐸𝐸(�𝒙𝒙′, 𝑠𝑠) �𝒙𝒙′ −∇𝐸𝐸( �𝒙𝒙′, 𝑠𝑠′) 𝐷𝐷 Figure 1: Energy-based interpretation for fluency degra- dation. Deceiving energy-based discriminator D re- quires 1) minimizing the energy E between the trans- ferred sentence �x′ and the target style s′, and 2) maximiz- ing the energy between the sentence �x′ and the original style s. However, the style s and s′ are originated from the same corpus, so maximizing E(�x′, s) degrades the overall fluency. It is an interpretation of Eq. 6. transfer approaches have become popular, includ- ing those using an autoencoder (Hu et al., 2017; Huang et al., 2020), back-translation (Prabhumoye et al., 2018; Lample et al., 2019), and reinforce- ment learning (Xu et al., 2018; Luo et al., 2019). Among previous studies, Style Transformer (Dai et al., 2019) achieved fine-grained style control by deceiving the style discriminator through adversar- …

Content truncated to 500 words