Detecting Bot-Generated Text by Characterizing Linguistic Accommodation in Human-Bot Interactions

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3235–3247 August 1–6, 2021. ©2021 Association for Computational Linguistics 3235 Detecting Bot-Generated Text by Characterizing Linguistic Accommodation in Human-Bot Interactions Paras Bhatt and Anthony Rios Department of Information Systems and Cyber Security University of Texas at San Antonio {Paras.Bhatt, Anthony.Rios}@utsa.edu Abstract Language generation models’ democratiza- tion benefits many domains, from answer- ing health-related questions to enhancing ed- ucation by providing AI-driven tutoring ser- vices. However, language generation mod- els’ democratization also makes it easier to generate human-like text at-scale for nefari- ous activities, from spreading misinformation to targeting specific groups with hate speech. Thus, it is essential to understand how peo- ple interact with bots and develop methods to detect bot-generated text. This paper shows that bot-generated text detection methods are more robust across datasets and models if we use information about how people respond to it rather than using the bot’s text directly. We also analyze linguistic alignment, provid- ing insight into differences between human- human and human-bot conversations. 1 Introduction Bots are useful in a wide variety of appli- cations areas including business (Kaczorowska- Spychalska, 2019), education (Kerlyl et al., 2006), and health (Yadav et al., 2019; Liednikova et al., 2020). For instance, Yadav et al. (2019) stud- ied the use of chatbots as a drop-in first-point-of- contact for women in India seeking breastfeeding information. Similarly, researchers have studied bots to answer COVID-19-related questions (Oni- ani and Wang, 2020) and screen individuals for risks of contracting the virus (Martin et al., 2020). Overall, the wide availability of software pack- ages, tools, and pre-trained models has democra- tized the creation of bots. Even with the increasing interest in bots for so- cial good (e.g., COVID-related chatbots), there is still a concern regarding their abuse to spread mis- information, be used for targeted discrimination, deceive users, and perform fraud (Daniel et al., 2019). Given the potential good and harm bots can create, it is essential to study how the bots should act and how people do interact with such bots for specific applications. Automatically identifying bots online is well studied (Garcia-Silva et al., 2019; Herzig et al., 2019; Kosmajac and Keselj, 2019; Ippolito et al., 2020; Jawahar et al., 2020). Bot detection meth- ods rely on two forms of information (Orabi et al., 2020): behavior and content. Behavior relates to measuring how often bots post, the time posts are created, and conversational net- work structures (Beskow and Carley, 2018). Con- tent involves using the bot’s text directly. Fo- cusing on text suffers from generalization issues, making it challenging to detect bots that discuss different topics or operate in different domains. Likewise, behavioral approaches assume that bots will behave differently than people at a superfi- cial level, such as posting more often than hu- mans. Even in the network analysis of conversa- tion structures (Beskow and Carley, 2018), many of the human-bot interactions on social media are not human-like interactions. For instance, it is understandable that humans will interact with a …

Content truncated to 500 words