首页|COSYWA: Enhancing Semantic Integrity in Watermarking Natural Language Generation
COSYWA: Enhancing Semantic Integrity in Watermarking Natural Language Generation
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
With the increasing use of natural language generation (NLG) models， there is a growing need to differentiate between machine-generated text and natural language text。 One promising approach is watermarking， which can help identify machine-generated text and protect against risks such as spam emails and academic dishonesty。 However， existing watermarking methods can significantly affect the semantic meaning of the text， creating a need for more effective techniques that maintain semantic integrity。 In this paper， we propose a novel watermarking method called Contextual SYnonym WAtermarking (COSYWA) that embeds watermarks in text using a Masked Language Model (MLM) without significantly impairing its semantics。 Specifically， we use postprocessing to embed watermarks in the output of an NLG model。 We generate a context-based synonym set using an MLM model to embed watermark information and use statistical hypothesis testing to detect the existence of watermarking。 Our experimental results show that COSYWA significantly enhances the text's capacity to maintain its original meaning while effectively embedding a watermark， making it a promising approach for protecting against misinformation in NLG。
Natural Language GenerationWatermarkingContextual Synonym
Junjie Fang、Zhixing Tan、Xiaodong Shi
展开 >
Department of Artificial Intelligence, School of Informatics, Xiamen University,Xiamen, China
Zhongguancun Laboratory, Beijing, People's Republic of China
Department of Artificial Intelligence, School of Informatics, Xiamen University,Xiamen, China##Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage of Pujian and Taiwan (Xiamen University), Ministry of Culture and Tourism, Xiamen, China
International conference on natural language processing and Chinese computing
NETL
