Looking for Traces of Textual Deepfakes in Bulgarian on Social Media

Proceedings of Recent Advances in Natural Language Processing, pages 1151–1161 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_122 1151 Looking for Traces of Textual Deepfakes in Bulgarian on Social Media Irina Temnikova1 Iva Marinova2 Silvia Gargova1 Ruslana Margova1 Ivan Koychev3 1Big Data for Smart Society Institute (GATE), Bulgaria, 2Institute of Information and Communication Technologies, Bulgarian Academy of Sciences, Bulgaria, 3Sofia University “St. Kliment Ohridski”, Bulgaria {irina.temnikova, silvia.gargova,ruslana.margova}@gate-ai.eu iva.marinova@identrics.ai koychev@fmi.uni-sofia.bg Abstract Textual deepfakes can cause harm, especially on social media. At the moment, there are mod- els trained to detect deepfake messages mainly for the English language, but no research or datasets currently exist for detecting them in most low-resource languages, such as Bulgar- ian. To address this gap, we explore three approaches. First, we machine translate an English-language social media dataset with bot messages into Bulgarian. However, the trans- lation quality is unsatisfactory, leading us to create a new Bulgarian-language dataset with real social media messages and those generated by two language models (a new Bulgarian GPT- 2 model – GPT-WEB-BG 1, and ChatGPT). We machine translate it into English and test exist- ing English GPT-2 and ChatGPT detectors on it, achieving only 0.44-0.51 accuracy. Next, we train our own classifiers on the Bulgarian dataset, obtaining an accuracy of 0.97. Addi- tionally, we apply the classifier with the highest results to a recently released Bulgarian social media dataset with manually fact-checked mes- sages, which successfully identifies some of the messages as generated by Language Models (LM). Our results show that the use of machine translation is not suitable for textual deepfakes detection. We conclude that combining LM text detection with fact-checking is the most appropriate method for this task, and that iden- tifying Bulgarian textual deepfakes is indeed possible. 1 Introduction The term “deepfake”, comes from “deep learning” and “fake” and indicates (potentially) fake texts, images, or videos, generated using deep learning models (Gambini, 2020). Among them, “Textual DeepFakes” (TDF) refer to texts generated automat- ically with the help of Generative Models (GMs, 1https://huggingface.co/usmiva/ gpt-web-bg. and lately with Large Language Models - LLMs), which may also contain fake or untrue content. This makes those of them, which are spread with the in- tention to deceive, the automatic variant of disinfor- mation (as defined by the European Commission (EC)2. According to this EC’s definition, “disin- formation” is “false or misleading content that is spread with an intention to deceive or secure eco- nomic or political gain and which may cause public harm”. There exist useful GMs and LLMs appli- cations (Kasneci et al., 2023). However, textual deepfakes can be a serious problem when spread on official information channels, as they can reach a large number of people. They are also problematic because, when fluent, they are hard to recognize by humans (Crothers et al., 2022). TDFs can be used by politicians in their political fights and destroy a person’s reputation, or to influence a large num- ber of people about sensitive topics such as a war or health. TDFs can be especially problematic on …

Content truncated to 500 words