A semantics-based watermarking framework, SemaMark, is proposed to enhance the robustness of LLM-generated text detection against paraphrasing.