Uncertainty Detection for Natural Language Watermarking

International Joint Conference on Natural Language Processing, pages 1188–1194, Nagoya, Japan, 14-18 October 2013. Uncertainty Detection for Natural Language Watermarking Gy¨orgy Szarvas1 Iryna Gurevych1,2 (1) Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt (2) Ubiquitous Knowledge Processing Lab (UKP-DIPF) German Institute for Educational Research and Educational Information http://www.ukp.tu-darmstadt.de Abstract In this paper we investigate the applica- tion of uncertainty detection to text wa- termarking, a problem where the aim is to produce individually identifiable copies of a source text via small manipulations to the text (e.g. synonym substitutions). As previous attempts showed, accurate para- phrasing is challenging in an open vocab- ulary setting, so we propose the use of the closed word class of uncertainty cues. We demonstrate that these words are promis- ing for text watermarking as they can be accurately disambiguated (from the non- cue uses of the same words) and their sub- stitution with other cues has marginal im- pact to the meaning of the text. 1 Introduction The goal of digital watermarking is to hide digital information (a secret marker) in an audio stream, image or text file. These markers are by design not perceivable while listening, watching or read- ing the data, but can be read with a tailor-made algorithm and can be used to authenticate the data that carries it, or to identify its owner. We discuss the concept of a watermark and the process of em- bedding it in a media in more detail in Section 2. Text watermarking is a digital watermarking prob- lem where the aim is to embed a secret message (a sequence of bits) in a text in order to make the ac- tual text copy individually identifiable (Bergmair, 2007). That is, given a natural language source text (e.g. an ebook), the aim is to produce individ- ual copies of the text by means of surface, syn- tactic or semantic manipulations. The individual- ized copies should preserve the readability and the meaning of the original text, i.e. the modifications should be undetectable to the readers. The manipulation of the text can be performed on the surface or the content level. Surface manip- ulation affects the visual appearance of the text, as in white space modulation. In contrast, syntactic reordering looks e.g. for conjunctions and reorders the connected words or phrases. Finally, semantic manipulation, such as synonym substitution, takes a target word and replaces it with a contextually plausible synonym. Consider the following exam- ples: He was bright and independent and proud. He was bright and independent and proud. (surface) He was independent and bright and proud. (syntactic) He was clever and independent and proud. (semantic) In steganography, there is a natural tradeoff be- tween capacity (the length of the secret message that can be embedded in a cover medium) and precision (how well the manipulations preserve readability and meaning). Text watermarking is a precision- oriented application as it requires very high transformation quality, and relatively low ca- pacity (recall) is acceptable as the goal is …

Content truncated to 500 words