TripPattern improves text watermarks without hurting quality
TripPattern: A Pattern-based Text Watermarking Method for Large Language Models
Artificial Intelligence
Summary
Large language models can produce text that might be hard to tell apart from human writing, so people create watermarks to detect machine-generated text. Existing watermark methods can make the text sound less natural because they push the model toward certain words. The authors propose TripPattern, which splits words into three groups and uses patterns between two groups while allowing neutral words freely. This keeps the text natural while embedding detectable patterns. Their tests show TripPattern keeps text quality good and still finds machine-generated text reliably.
What this means in practice
- •For content moderation teams: Identify machine-generated text in multilingual online content without sacrificing text naturalness.
- •For ai service operators: Embed detectable watermarks in outputs from large language models to ensure traceability while maintaining text quality.$Commercial implications: Enables watermarking features integrated into AI-generated content services for compliance and authenticity verification.
Authors
Sangjun Moon, Dasom Choi, Jingun Kwon, Hidetaka Kamigaito, Taro Watanabe, Manabu Okumura
Abstract
Text watermarking techniques have gained significant attention for identifying machine-generated text and mitigating risks from large language models (LLMs). Existing methods typically divide an LLM's vocabulary into green and red tokens, but encouraging generation toward green tokens can reduce text quality and naturalness. To address this, we propose TripPattern, a watermarking framework that formulates text watermarking as a pattern-based matching task using three vocabulary partitions. TripPattern divides the vocabulary into one neutral group and two pattern groups. During generation, the model alternates token selection between the two pattern groups to embed detectable patterns, while neutral tokens are selected independently to improve flexibility and preserve naturalness. For detection, TripPattern uses pattern-based statistical tests that provide interpretable p-values by measuring how often adjacent tokens alternate between the pattern groups. Theoretical analysis and empirical evaluations on four multilingual datasets show that TripPattern maintains LLM generation quality while achieving robust watermark detectability.