Cross-Domain Detection of GPT-2-Generated Technical Text

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1213 - 1233 July 10-15, 2022 ©2022 Association for Computational Linguistics Cross-Domain Detection of GPT-2-Generated Technical Text Juan Diego Rodriguez1 Todd Hay2 David Gros3 Zain Shamsi1 Ravi Srinivasan1 1Applied Research Laboratories, The University of Texas at Austin 2En Solución 3University of California, Davis {juand-r, zain.shamsi, ravi.srinivasan}@utexas.edu todd.hay@ensolucion.com dgros@ucdavis.edu Abstract Machine-generated text presents a potential threat not only to the public sphere, but also to the scientific enterprise, whereby genuine re- search is undermined by convincing, synthetic text. In this paper we examine the problem of detecting GPT-2-generated technical research text. We first consider the realistic scenario where the defender does not have full infor- mation about the adversary’s text generation pipeline, but is able to label small amounts of in-domain genuine and synthetic text in or- der to adapt to the target distribution. Even in the extreme scenario of adapting a physics- domain detector to a biomedical detector, we find that only a few hundred labels are suffi- cient for good performance. Finally, we show that paragraph-level detectors can be used to detect the tampering of full-length documents under a variety of threat models. 1 Introduction Recent advances in techniques for generating real- istic synthetic content (i.e., deepfakes) pose a di- verse set of problems with significant societal con- sequences (Kreps et al., 2020; Bommasani et al., 2021; McGuffie and Newhouse, 2020). The advent of large language models for text generation (Rad- ford et al., 2019; Brown et al., 2020) have made it easier than ever to create convincing synthetic1 text (Solaiman et al., 2019). While much attention has been focused on the role of synthetic audio and video, it can be argued that synthetic text may give rise to some of the most serious threats to infor- mation integrity and the long-term preservation of archival knowledge (Aliman and Kester, 2021). 1Some authors use the term fake text; however, we prefer synthetic, since human-written text can contain false infor- mation, while machine-generated text can (sometimes) be factually correct (Schuster et al., 2020; Belz, 2019). Figure 1: Machine-generated text enables the corrup- tion of technical knowledge (e.g., biomedical research). One of the above abstracts was generated by GPT-2. More examples of generated text can be found in §A.7. While there are currently no documented cases of published papers containing text from neural lan- guage models, non-neural machine-generated pa- pers have already been published in peer-reviewed journals (Cabanac and Labbé, 2021). With more convincing text generation techniques and a grow- ing number of publications, this is a problem that could get much worse, particularly for non-peer- reviewed technical text (Ranade et al., 2021). In ad- dition to hindering the scientific process, synthetic technical texts could be used to manipulate public opinion and sow discord around specific scientific topics (Aliman and Kester, 2021). Other conse- quences of machine-generated technical text in- clude the contamination of NLP pipelines (Ranade et al., 2021) and poisoning …

Content truncated to 500 words