Text watermarking has emerged as an important technique for detecting machine-generated text. However, existing methods generally use arbitrary vocabulary partitioning during decoding, which results in the absence of appropriate words during the response generation and disrupts the language model's expressiveness, thus severely degrading the quality of text response. To address these issues, we introduce a novel approach, Watermarking with Mutual Exclusion (WatME). Specifically, by leveraging linguistic prior knowledge of inherent lexical redundancy, WatME can dynamically optimize the use of available vocabulary during the decoding process of language models. It employs a mutually exclusive rule to manage this redundancy, avoiding situations where appropriate words are unavailable and maintaining the expressive power of large language models (LLMs). We present theoretical analysis and empirical evidence demonstrating that WatME substantially preserves the text generation ability of LLMs while maintaining watermark detectability. Specifically, we investigate watermarking's impact on the emergent abilities of LLMs, including knowledge recall and logical reasoning. Our comprehensive experiments confirm that WatME consistently outperforms existing methods in retaining these crucial capabilities of LLMs. Our code will be released to facilitate future research.

翻译：摘要：文本水印技术已成为检测机器生成文本的重要手段。然而现有方法通常在解码过程中采用任意词汇划分策略，导致响应生成时缺乏恰当词汇，破坏语言模型的表达能力，从而显著降低文本响应质量。为解决这些问题，我们提出了一种新颖方法——互斥水印技术（WatME）。具体而言，WatME通过利用词汇固有冗余的语言学先验知识，可在语言模型解码过程中动态优化可用词汇的使用。该方法采用互斥规则管理词汇冗余，避免出现恰当词汇不可用的情况，同时保持大语言模型（LLM）的表达能力。我们通过理论分析和实验证据表明，WatME在保持水印可检测性的前提下，显著保留了LLM的文本生成能力。特别地，我们研究了水印技术对LLM涌现能力（包括知识回忆与逻辑推理）的影响。全面实验证实，WatME在保留LLM这些关键能力方面始终优于现有方法。我们将公开代码以促进后续研究。