Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Steering Large Language Models with Register Analysis for Arbitrary Style Transfer

Xinchen Yang, Marine Carpuat

TL;DR

The paper addresses arbitrary style transfer using LLMs by proposing a register-analysis–driven prompting framework. It introduces two prompting variants, RG-Contrastive and RG, which leverage Biber's multidimensional register analysis to generate target style descriptors guiding rewrites of input texts $x^{\mathrm{input}}$ toward exemplars $x^{\mathrm{style}}$, producing outputs $x^{\mathrm{output}}$. Across Authorship Imitation, Formality Transfer, and Text Simplification, the approach yields improved meaning preservation and competitive style transfer strength, outperforming several baselines and STYLL in key tasks. Qualitative analysis confirms descriptors align with register in a way that minimizes semantic drift, underscoring better decoupling of style and content and supporting practical use in low-resource settings with responsible deployment considerations.

Abstract

Large Language Models (LLMs) have demonstrated strong capabilities in rewriting text across various styles. However, effectively leveraging this ability for example-based arbitrary style transfer, where an input text is rewritten to match the style of a given exemplar, remains an open challenge. A key question is how to describe the style of the exemplar to guide LLMs toward high-quality rewrites. In this work, we propose a prompting method based on register analysis to guide LLMs to perform this task. Empirical evaluations across multiple style transfer tasks show that our prompting approach enhances style transfer strength while preserving meaning more effectively than existing prompting strategies.

Steering Large Language Models with Register Analysis for Arbitrary Style Transfer

TL;DR

The paper addresses arbitrary style transfer using LLMs by proposing a register-analysis–driven prompting framework. It introduces two prompting variants, RG-Contrastive and RG, which leverage Biber's multidimensional register analysis to generate target style descriptors guiding rewrites of input texts toward exemplars , producing outputs . Across Authorship Imitation, Formality Transfer, and Text Simplification, the approach yields improved meaning preservation and competitive style transfer strength, outperforming several baselines and STYLL in key tasks. Qualitative analysis confirms descriptors align with register in a way that minimizes semantic drift, underscoring better decoupling of style and content and supporting practical use in low-resource settings with responsible deployment considerations.

Abstract

Large Language Models (LLMs) have demonstrated strong capabilities in rewriting text across various styles. However, effectively leveraging this ability for example-based arbitrary style transfer, where an input text is rewritten to match the style of a given exemplar, remains an open challenge. A key question is how to describe the style of the exemplar to guide LLMs toward high-quality rewrites. In this work, we propose a prompting method based on register analysis to guide LLMs to perform this task. Empirical evaluations across multiple style transfer tasks show that our prompting approach enhances style transfer strength while preserving meaning more effectively than existing prompting strategies.
Paper Structure (44 sections, 2 equations, 4 figures, 9 tables)

This paper contains 44 sections, 2 equations, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Style and Register in Corpus Linguistics
  4. Style Transfer in NLP
  5. Style in NLP.
  6. Supervised Single-Style Transfer
  7. Multitask Style Transfer.
  8. LLM-Based Style Transfer.
  9. Approach
  10. Experiment Setup
  11. Tasks
  12. Authorship Imitation
  13. Datasets.
  14. Task-based metrics.
  15. Formality Transfer
  16. ...and 29 more sections

Figures (4)

  • Figure 1: Prompting pipeline for RG-Contrastive and RG, respectively.
  • Figure 2: Style–meaning trade-offs across tasks: MUD, Cochrane, and GYAFC (both directions), using Llama-3.2-3B-Instruct. Pareto frontiers identify systems that achieve optimal trade-offs for each task. RG-C: RG-Contrastive.
  • Figure 3: Top 15 style descriptors by frequency by generated rewriting system (RG-Contrastive, RG, STYLL) on the MUD_Random task.
  • Figure 4: Illustration of the procedure of building Biber’s MDA representation from a linguistic corpus and using it to make inference on a new text following the practice of jackgrieve.