SemStyle: A Robust Sentence-Level Watermarking Framework Through Semantic-to-Style Mapping

The proliferation of large language models (LLMs) necessitates effective text watermarking for copyright protection and misinformation tracing. However, existing token-level methods are inherently vulnerable to paraphrasing attacks, while recent sentence-level alternatives suffer from severe computational overhead due to rejection sampling. To address these limitations, we propose SemStyle, a sentence-level watermarking framework based on semantic-to-style mapping. Specifically, SemStyle extracts semantic contexts from preceding sentences to generate style fingerprint sequences, which are then used to guide the linguistic style of generated sentences. Instead of unlimited re-sampling, SemStyle selects the generated sentence from a fixed-size set of candidates that best matches the target style. Experiments on three datasets show that SemStyle maintains generation quality comparable to existing baselines while achieving improved detection robustness under the evaluated paraphrasing and back-translation settings. Furthermore, SemStyle achieves a generation speed over $2\times$ higher than existing sentence-level watermarking mechanisms.