A reversible natural language watermarking for sensitive information protection

Lingyun Xiang, Yangfan Liu, Zhongliang Yang

Abstract: Highlights•The first natural language watermarking method for sensitive information protection.•A keyword extraction algorithm is employed to identify sensitive information.•A masked language model is utilized to predict appropriate substitution words.•A prediction error expansion algorithm is employed for reversible embedding.•KSPEE outperforms baselines in text quality and exhibits high security and bpw.