SC2: Towards Enhancing Content Preservation and Style Consistency in Long Text Style Transfer

Text style transfer (TST) aims to vary the style polarity of text whilepreserving the semantic content. Although recent advancements have demonstratedremarkable progress in short TST, it remains a relatively straightforward taskwith limited practical applications. The more comprehensive long TST taskpresents two challenges: (1) existing methods encounter difficulties inaccurately evaluating content attributes in multiple words, leading to contentdegradation; (2) the conventional vanilla style classifier loss encountersobstacles in maintaining consistent style across multiple generated sentences. In this paper, we propose a novel method SC2, where a multilayer JointStyle-Content Weighed (JSCW) module and a Style Consistency loss are designedto address the two issues. The JSCW simultaneously assesses the amounts ofstyle and content attributes within a token, aiming to acquire a losslesscontent representation and thereby enhancing content preservation. The multipleJSCW layers further progressively refine content representations. We design astyle consistency loss to ensure the generated multiple sentences consistentlyreflect the target style polarity. Moreover, we incorporate a denoisingnon-autoregressive decoder to accelerate the training. We conduct plentifulexperiments and the results show significant improvements of SC2 overcompetitive baselines. Our code: https://github.com/jiezhao6/SC2.