Semantic-Unit Order Watermarking for AI-Generated Text: A Constrained Discourse-Functional Framework for Provenance Encoding and Verification
The rapid development of large language models (LLMs) has increased the difficulty of establishing the provenance of generated text, particularly when outputs undergo paraphrasing, rewriting, restructuring, and human editing. Existing text-watermarking approaches have primarily focused on token-selection distributions, lexical substitutions, syntactic transformations, character-level modifications, semantic representations, and discourse-level relationships. However, the ordering of explicit semantic-functional units—such as definitions, explanations, arguments, evidence, examples, counterarguments, and conclusions—provides a distinct potential encoding dimension that warrants systematic investigation. This paper introduces Semantic-Unit Order Watermarking (SUOW), a proposed framework that encodes watermark information through constrained ordering patterns of semantic-functional units in AI-generated text. Unlike approaches that treat individual words, tokens, or sentences as the primary encoding units, SUOW operates on functionally characterized segments whose relationships contribute to the organization of an argument or explanation. Its central principle is to exploit permissible alternative orderings while preserving essential semantic content, rhetorical coherence, logical dependencies, and discourse completeness. The framework introduces a semantic-unit representation model, a dependency-constrained ordering space, a key-controlled permutation encoder, a watermark embedding mechanism, a blind detection procedure, and a provenance verification model. A directed acyclic graph is employed to represent mandatory dependencies among semantic units, allowing candidate orderings to be generated through valid topological arrangements rather than unrestricted permutations. The theoretical encoding capacity is consequently determined by the number of admissible semantic-unit orderings. The paper further develops a formal optimization objective and an experimental protocol for evaluating detection accuracy, semantic preservation, argument coherence, robustness against paraphrasing and structural rewriting, and false-positive behavior. To the best of our knowledge, this paper is the first to formulate a dedicated watermarking framework centered specifically on the constrained ordering of explicitly classified semantic-functional units as the primary carrier of watermark information in AI-generated text. This claim is distinguished from earlier semantic and discourse-level watermarking research, which is reviewed and acknowledged. The proposed framework is theoretical and methodological; its practical effectiveness requires empirical validation.
Authors
- Jincheng Zhang (ORCID: https://orcid.org/0009-0005-1860-0009)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24392183
- Primary Topic
- Advanced Steganography and Watermarking Techniques
- Type
- preprint