自然语言水印隐蔽性自动评测方法

Automatic evaluation scheme for imperceptibility of natural language watermarking

对自然语言水印(NLW)隐蔽性评测的缺乏严重影响了该领域技术的发展,为此,结合NLW的特点和语言心理学,以人类对语言释义的速度及难易程度为依据,提出了NLW隐蔽性的感知模型和相应的评测方案:从语法错误、搭配错误和语义损失三个方面对隐蔽性进行评测。最后,针对四种不同水印技术(基于绝对同义词替换的TLex水印系统、相对同义词替换水印系统、基于句法树的水印系统和中文句法水印系统),一方面利用该方案进行自动评测,另一方面进行置信度为90%的人工评测。两种方法得到了相同的结论:基于词汇的NLW技术的隐蔽性优于基于句子的NLW技术,说明该自动评测方法是评测NLW隐蔽性的有效评测方法。

The deficiency of evaluation scheme for the imperceptibility of Natural Language Watermarking （NLW） seriously restricts the development in this field. To resolve this problem, combining the characteristics of natural language and the knowledge of psycholinguistics and considering the difficulty for human to understand natural language, an imperceptibility perceptual model and an automatic evaluation scheme were proposed. The proposed scheme evaluated the imperceptibility from three aspects： syntax error, collocation error and semantic loss. Finally, four different watermarking schemes including the T- Lex watermarking system based on absolute synonym substitution, the relative synonymous substitution watermarking system, the watermarking system based on syntactic tree and the Chinese syntactic watermarking system were used to compare the results of the proposed automatic method and manual method whose confidence coefficient is 90%. The two methods get the same conclusion： the imperceptibility of NLW based on word is better than that of NLW based on sentence, and the automatic evaluation scheme is effective for the imperceptibility of NLW.