自然語言生成的技術已經可以生成通順流暢的文章，但也造成了一些潛在的風險，這項技術可能會被惡意使用來生成文章誤導大眾。為了避免這些風險，需要發展辨識機器生成文章的技術來偵測這些文章。然而由於人類難以辨識使用自然語言生成技術產生的文章，在現實環境下要收集來自同一攻擊者的機器文章是很困難的，這也造成我們難以使用充足的資料訓練辨識器。即使我們想自行訓練生成模型，並用來增加機器類別的資料量，我們也無法得知攻擊者是如何生成資料的，因此難以仿造攻擊者的生成方法產生更多資料。

我們的論文研究了不同生成方法之間的區別，並且發現辨識器難以辨識藉由與訓練集不同方法生成的文章，這也表示若使用差異過大之生成方法產生更多機器類別的資料，不但無法改進辨識器對收集到之機器資料的辨識能力，還可能使效果更差。因此我們提出了一種資料增強方法並取名為Copycat，Copycat會藉由計算顯著性分數來得到關於原始機器資料生成方式的資訊，藉此模仿原始機器類別資料之生成方式產生額外的機器資料，我們的實驗結果表明藉由Copycat生成的額外資料可以更好的改善辨識器的效果，表現優於其他任務無關的資料增強方法以及使用任意生成方式增加資料量的做法。

Recent developments in natural language generation have made it possible to generate fluent articles automatically, but it also poses some potential risks. This technology may be used maliciously to generate articles that can mislead the public. To avoid these risks, building automatic discriminators for detecting machine-generated texts is required. However, because humans are poor at identifying articles generated by natural language generation techniques, machine articles from the same adversary are hard to collect in real-world situations. It is hard to train a discriminator with sufficient data. Even if we want to train the generator by ourselves and use it to increase the amount of machine data, we cannot know how the adversary generates the article, so it is difficult to imitate the adversary's generation approach to generate more machine data.
Our paper investigates the differences between different generation approaches. We find discriminators perform poorly in detecting articles generated by a generation approach different from the training set. This means that using a generation approach different from the original machine data will be ineffective in improving the discriminator's performance. Through the saliency score technology, Copycat obtains information about the generation approach of machine data. With this information, Copycat can mimic the generation approach of machine data to generate synthetic machine data. Our experimental results show that the synthetic data generated by Copycat can better improve the discriminator's performance and outperforms other data augmentation methods.

[1] Ellie Pavlick Stefanie Tellex Aaron Gokaslan*, Vanya Cohen*. Openwebtext corpus, 2019.
[2] David Ifeoluwa Adelani, Haotian Mai, Fuming Fang, Huy H Nguyen, Junichi Yamagishi, and Isao Echizen. Generating sentiment­preserving fake online reviews using neural language models and their human­and machine­based detection. In International Conference on Advanced Information Networking and Applications, pages 1341–1354. Springer, 2020.
[3] Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc’Aurelio Ranzato, and Arthur Szlam. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351, 2019.
[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert­Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few­shot learners. 2020.
[5] Jacob Devlin, Ming­Wei Chang, and Kenton Lee. Google, kt, language, ai: Bert: pretraining of deep bidirectional transformers for language understanding. In Proceedings of NAACL­HLT, pages 4171–4186, 2018.
[6] Chris Donahue, Mina Lee, and Percy Liang. Enabling language models to fill in the blanks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2492–2501, 2020.
[7] Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. Tweepfake: About detecting deepfake tweets. Plos one, 16(5):e0251415, 2021.
[8] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 889–898, 2018.
[9] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of generated text. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 111– 116, 2019.
[10] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations, 2019. 46
[11] Fei Huang, Zikai Chen, Chen Henry Wu, Qihan Guo, Xiaoyan Zhu, and Minlie Huang. Nast: A non­autoregressive generator with word alignment for unsupervised text style transfer. In Findings of the Association for Computational Linguistics: ACL­IJCNLP 2021, pages 1577–1590, 2021.
[12] Daphne Ippolito, Daniel Duckworth, Chris Callison­Burch, and Douglas Eck. Automatic detection of generated text is easiest when humans are fooled. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1808–1822, 2020.
[13] Ganesh Jawahar, Muhammad Abdul­Mageed, and VS Laks Lakshmanan. Automatic detection of machine generated text: A critical survey. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2296–2309, 2020.
[14] Nitish Shirish Keskar, Bryan McCann, Lav Varshney, Caiming Xiong, and Richard Socher. CTRL ­ A Conditional Transformer Language Model for Controllable Generation. arXiv preprint arXiv:1909.05858, 2019.
[15] Tharindu Kumarage, Amrita Bhattacharjee, Kai Shu, and Huan Liu. Data generation for neural disinformation detection.
[16] Jiwei Li, Xinlei Chen, Eduard Hovy, and Dan Jurafsky. Visualizing and understanding neural models in nlp. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 681–691, 2016.
[17] Dayiheng Liu, Jie Fu, Pengfei Liu, and Jiancheng Lv. Tigs: An inference algorithm for text infilling with gradient search. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4146–4156, 2019.
[18] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.
[19] Shayne Longpre, Yu Wang, and Chris DuBois. How effective is task­agnostic data augmentation for pretrained transformers? In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4401–4411, 2020.
[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018.
[21] J. Mackenzie, R. Benham, M. Petri, J. R. Trippas, J. S. Culpepper, and A. Moffat. Ccnews­en: A large english news corpus. In Proc. CIKM, pages 3077–3084, 2020.
[22] Eric Malmi, Aliaksei Severyn, and Sascha Rothe. Unsupervised text style transfer with padded masked language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8671–8680, 2020.
[23] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016.
[24] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.47
[25] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprintarXiv:1312.6034, 2013.
[26] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert­Voss, JeffWu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Releasestrategies and the social impacts of language models. arXiv preprint arXiv:1908.09203,2019.
[27] Jason Wei and Kai Zou. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP­IJCNLP), pages 6382–6388, 2019.
[28] Xing Wu, Shangwen Lv, Liangjun Zang, Jizhong Han, and Songlin Hu. Conditionalbert contextual augmentation. In International Conference on Computational Science, pages 84–95. Springer, 2019.
[29] Jingjing Xu, Xu Sun, Qi Zeng, Xuancheng Ren, Xiaodong Zhang, Houfeng Wang, andWenjie Li. Unpaired sentiment­to­sentiment translation: A cycled reinforcement learning approach. In ACL, 2018.
[30] Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, FranziskaRoesner, and Yejin Choi. Defending against neural fake news. Advances in neural information processing systems, 32, 2019.
[31] Xiang Zhang, Junbo Zhao, and Yann LeCun. Character­level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015.
[32] Yizhe Zhang, Siqi Sun, Michel Galley, Yen­Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B Dolan. Dialogpt: Large­scale generativepre­training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270–278, 2020.
[33] Wangchunshu Zhou, Tao Ge, Canwen Xu, Ke Xu, and Furu Wei. Improving sequence to ­sequence pre­training via sequence span rewriting. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 571–582, Online and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics.
[34] Wanrong Zhu, Zhiting Hu, and Eric Xing. Text infilling. arXiv preprintarXiv:1901.00158, 2019.
[35] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story­like visual explanations by watching movies and reading books. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 19–27. IEEE, 2015