Beyond Simple Text Style Transfer: Unveiling Compound Text Style Transfer with Prompt-Based Pre-Trained Language Models

Abstract

Compound text style transfer is an innovative task that seeks to merge textual elements from distinct styles, themes, or attributes to create diverse and distinctive textual content. This technique plays an important role in various fields, such as personalized storyline generation for game characters and movie plot development. However, it faces certain challenges, including model size limitations, data scarcity, and evaluation difficulties. To address these challenges, we present corresponding solutions. First, we leverage GPT-3.5-turbo to build a comprehensive framework for generating compound style corpora to address data scarcity. Second, we propose Multi-Prompts-Fusion-Framework (MPFF), a data-centric compound text style transfer framework to compensate for the impact of model size reduction. Finally, we use a GPT-3.5-turbo-based evaluation method to assess semantic preservation, style transfer consistency, and grammatical fluency. Our experiments consistently demonstrate the effectiveness of our compound text style transfer framework. Code and data are available at https://github.com/Arthas183/MPFF.