Style transfer for text is the task of rewriting an input text to exhibit a particular style while maintaining the text's overall meaning. For example, 'It was a lovely day' could be rewritten in a 'more flowery' style: 'It was a day as lovely as a spring rainshower'. Current models achieve good performance on styles for which labeled training data is available, such as changing the sentiment of a movie review from positive to negative. There have recently been a few exciting new label-free approaches, which do not require a fully-labeled corpus. Following this new trend, we explore style transfer using large language models (LLMs), which enable label-free approaches to NLP tasks and achieve strong performance in a zero-shot setting. More importantly, LLMs fundamentally change the landscape of possible style transformations: they are capable of arbitrary transformations that can be described in plain english (e.g., "make this sentence more melodramatic", "include the word 'avocado'", or "write this sentence in the style of Stephen King"). This ability to transform text in arbitrary ways unlocks a variety of real-world applications of NLP style transfer systems such as assisted writing, mitigating online toxicity, and facilitating personalized responses from chatbots

The augmented zero shot prompts used to generate the outputs for our proposed method are available for download here for the dialog model and here for GLM/GPT-3.

Here are the results from our method compared with some baselines. Our method, augmented zero shot, is shown in blue.