SWAN: Semantic Watermarking with Abstract Meaning Representation
Researchers have developed SWAN, a new framework for embedding watermarks into the semantic structure of sentences using Abstract Meaning Representation (AMR). Unlike previous methods that alter token selection, SWAN encodes signatures directly into the semantic representation, making them robust to paraphrasing. This training-free approach uses LLM prompting for injection and an AMR parser for detection, achieving state-of-the-art performance and improving detection AUC by up to 13.9 percentage points against paraphrasing. AI
IMPACT Introduces a more robust method for text provenance verification, potentially aiding in detecting AI-generated content even after modifications.