Text Watermarking Based on Syntactic Constituent Movement

Abstract

This paper explores a method of text watermarking for agglutinative languages and develops a syntactic tree-based syntactic constituent movement scheme. Agglutinative languages provide a good ground for the syntactic tree-based natural language watermarking because syntactic constituent order is relatively free. Our proposed natural language watermarking method consists of seven procedures. First, we construct a syntactic dependency tree of unmarked text. Next, we perform clausal segmentation from the syntactic tree. Third, we choose target syntactic constituents, which will move within its clause. Fourth, we determine the movement direction of the target constituents. Then, we embed a watermark bit for each target constituent. Sixth, if the watermark bit does not coincide with the direction of the target constituent movement, we displace the target constituent in the syntactic tree. Finally, from the modified syntactic tree, we obtain a marked text. From the experimental results, we show that the coverage of our method is 91.53%, and the rate of unnatural sentences of marked text is 23.16%, which is better than that of previous systems. Experimental results also show that the marked text keeps the same style, and it has the same information without semantic distortion.

Generating watermarked adversarial texts

“…Mainstream algorithms exploit the syntactic and semantic nature of texts for watermark embedding 43 – 46 They emphasize that the critical semantic information of the marked texts are consistent with the original ones. For example, secret bits can be embedded into a text by replacing specified words in the text with semantically similar words.…”