What Is This Book About?
We live in a world where social media is deeply embedded in many people’s lives, with some spending hours each day on platforms such as TikTok, Instagram, YouTube and Facebook. While these platforms have greatly expanded individuals’ opportunities for self-expression and engagement with diverse communities, a major challenge today is the widespread use of offensive language. Undoubtedly, the anonymity afforded by the internet has created a fertile environment in which people may feel less personally accountable and act more aggressively than they would offline. This has resulted in a range of experiences and conflicts, from feeling hurt or offended to more severe instances of online abuse and hate.
When people are interacting with others on social media, whether in their role as the speaker, hearer or recipient, or just a bystander, they are constantly and actively involved in the process of evaluating and interpreting one another’s language and behaviour. In this context, interactants tend to ‘present themselves in a particular way and align (or disalign) with others through the stances they take towards a particular idea, object or person’ (Tagg et al., Reference Tagg, Seargent and Brown2017, p. 44). This often leads to lengthy discussions over what counts as offensive and what does not, who has the right to be offended and who does not, and what taking offence might lead to. Sometimes, the subjective nature of this interpretation can create the impression that people take offence ‘too often and too easily’ (Barrow, Reference Barrow2005, p. 266). This is suggested by the frequent use of terms such as ‘snowflake generation’, which, according to the Collins English Dictionary (n.d.), refers to ‘the generation of people who became adults in or after the 2010s, viewed as being less resilient and more prone to taking offence than previous generations’. This situation has resulted in a rather complete failure to make a distinction between what is ‘merely offensive’ – yet protected under the right to freedom of expression – and what is grossly offensive or illegal, potentially warranting prosecution (cf. Bliss, Reference Bliss2017; Lepoutre, Reference Lepoutre2019). This distinction is by no means an easy subject to tackle; it is undeniably multifaceted and spans multiple disciplines.
Furthermore, online offensive language is often marked by a high degree of creativity, which poses a significant challenge for effective automatic identification and moderation. On the internet, users often adopt inventive and indirect strategies to stigmatise others, possibly with a view to bypassing detection mechanisms (Gröndahl et al., Reference Gröndahl, Pajola, Juuti, Conti and Asokan2018; Rawat et al., Reference Rawat, Kumar and Samant2024). However, relatively little research has been conducted on the creative ways language is used and shaped to express or provoke offence. This gap highlights the need for further exploration of how creativity may be exploited to offend others.
This book focuses specifically on online language identified as offensive by those directly targeted. The main question we aim to discuss is: What does offensive language online look like? To this end, we will explore its lexical, structural, discursive and pragmatic features.1 Accordingly, we will draw insights from multiple linguistic strands, primarily pragmatics, (im)politeness theories, discourse analysis and computational approaches to addressing online offence, as well as law and social psychology.
How Is This Book Different?
Offensive language has been at the forefront of various government reports globally. In the context of the UK, for example, a report by the Office for National Statistics (2023) reveals that one in five children aged 10 to 15 in England and Wales experienced at least one type of online bullying behaviour in the year ending March 2023. The need to moderate harmful and offensive online communication has been recognised by governments worldwide more than ever. Social media companies are consequently under increasing pressure to implement changes and update their policies and regulations to curb the spread of such language on the internet, as seen, for example, in the new Online Safety Act (2023) (see Chapter 1). In the majority of government reports, the most frequently identified form of online harassment involves being subjected to insults or receiving rude, offensive and hateful messages, though these categories are often presented without clear definitions or distinctions.
Online offensive language has also been investigated across various academic disciplines and is often discussed using broad terms such as cyberbullying, online abuse, harassment, hate speech and so on. This wide scope reflects the complexity of the issue, intersecting with human cognition and behaviour, societal norms and legal regulations. This multifaceted phenomenon has given rise to several important research agendas. The literature outside linguistics can be categorised into two main trends:
(a) Computational approaches and studies aim to develop, for example, algorithms to assist with the automatic detection and prevention of such language. These studies frequently use a ‘predictive modelling’ framework to create algorithms that facilitate the automated identification of offensive messages (e.g., Burnap & Williams, Reference Burnap and Williams2015; Gitari et al., Reference Gitari, Zuping, Damien and Long2015; Zhang et al., Reference Zhang, Robinson and Tepper2018). As discussed in Chapter 1, a key challenge in detecting online offensive language lies in current technological limitations. Many service providers continue to rely on manual checks to monitor offensive language (Abderrouaf & Oussalah, Reference Abderrouaf and Oussalah2019). Moreover, computational approaches designed to process and understand offensive and hateful language are only as effective and unbiased as the data on which they have been trained. These systems often reflect the biases, thought processes and moral frameworks of their creators, inevitably influencing their outputs. The reliance on significant human intervention at various stages of data training and hate detection demonstrates that this issue is fundamentally a linguistic challenge.
(b) Disciplines such as psychology, sociology, criminology and law, among others, engage with the study of online offence. In psychology, the focus is often on identifying psychological or personality traits commonly associated with victims and perpetrators of online offence (see Zhu et al., Reference Zhu, Huang, Evans and Zhang2021 for a review). Particular emphasis is placed on assessing ‘cybervictimisation’ experiences (Brochado et al., Reference Brochado, Soares and Fraga2017, p. 524) – that is, the experience of being targeted or harmed on digital platforms – and the psychological repercussions of such experiences across various demographics, including children, young and vulnerable adults (e.g., Keighley, Reference Keighley2022). Research in sociology and related fields, such as criminology, tends to examine the societal structures and cultural factors that enable or exacerbate more extreme forms of online behaviour, such as online hate. These studies explore what hate speech reveals about societal status and investigate the potential for online hate to provoke violence in offline contexts (see, e.g., Gorenc, Reference Gorenc2022; Matamoros-Fernández & Farkas, Reference Matamoros-Fernández and Farkas2021; Williams, Reference Williams2021). Legal studies, by contrast, focus on the application and suitability of legal frameworks and the prosecution of online behaviours (see, e.g., Askola, Reference Askola2015; Fino, Reference Fino2020; Solhjell, Reference Solhjell2023), including discussions on the boundaries of what constitutes hate speech (Brown & Sinclair, Reference Brown and Sinclair2023).
In most of the disciplines mentioned previously, existing research primarily focuses on the more severe end of the spectrum, examining the legal, social and psychological implications of hate speech and its regulation. However, less attention has been paid to language that, while not classified as hate speech, still has the potential to harm, offend or marginalise. This middle ground – language considered offensive but falling short of the legal threshold for hate speech – remains underexplored in academic research and government guidelines on online harm and hate. This is precisely the gap that this book seeks to fill. We are particularly concerned with uncovering the various layers of offensive language, including its lexical, semantic, structural and discursive-pragmatic features. As we will demonstrate, despite the wide variation in forms of expression, certain key features tend to recur, making this phenomenon well suited to linguistic analysis.
Furthermore, most studies proceed from the assumption that online victims, perpetrators, bystanders (i.e. those who witness online hate but are not its targets) and researchers all share a unanimous understanding of what constitutes offensive meaning. This assumption often rests on the common view that certain words and expressions inherently carry offensive meanings by virtue of their dictionary definitions (e.g., ‘You fucking faggot’). While it is undeniable that semantically offensive words are indeed used to offend, they represent only one of many ways in which hateful and offensive beliefs can be expressed. For instance, a comment such as, ‘You are dark and handsome. When it is dark, you are handsome.’ – which contains no overtly offensive words and even includes a positive adjective (‘handsome’) – could be equally, if not more, offensive. One of the major challenges in addressing offensive and hateful language is how to adequately account for context, without placing undue focus on linguistic forms to the detriment of situational, social and cultural dynamics, especially when conducting micro-level analysis (cf. Jucker et al., Reference Jucker, Schneider and Bublitz2018). It is against this backdrop that offensive language should be understood as identified by those who have experienced it, allowing the full context to be taken into account without speculation. What sets our study apart is its commitment to avoiding researcher bias and subjective interpretations of what counts as offensive. Instead, as we will explain, our analyses are grounded in naturally occurring language that has been genuinely perceived and flagged as offensive by those directly affected.
In addition, unlike prior research, this study does not focus solely on controversial events that exacerbate offensive language. Put simply, this study is not centred on ‘specific hate targets’ (ElSherief et al., Reference ElSherief, Ziems, Muchlinski, Anupindi, Seybolt, De Choudhury and Yang2021, p. 2), such as immigrants, asylum seekers or members of the LGBTQ+ community, who are often subjected to online hate and abuse. Although our data may include instances of, for example, racially targeted language, they are intended to be broad and to encompass a wider range of users.
The data for this study comprise examples of offensive language flagged as such by a group of internet users known as content creators or social media influencers (Khamis et al., Reference Khamis, Ang and Welling2017; Leban & Voyer, Reference Leban, Voyer, Yesiloglu and Costello2020). These individuals often start without prior fame and gradually build their online presence. They produce a wide range of content for their audiences, typically centred around everyday, routine activities, which contributes to their broad appeal, especially among ordinary internet users. Since both content creators and their followers are generally regarded as ordinary – meaning they are not politicians, macro-celebrities or professional activists – the comments they receive, which form the basis of our analysis, are likely to closely resemble the kinds of interactions most users encounter online. This is precisely why we believe this book makes a novel contribution to the literature, as the data used in this study offer a unique insight into offensive language as experienced and expressed by ordinary people.
Summary of the Chapters
In Chapter 1, we set the scene by examining the dynamics of online offensive language. We examine offensive language across a spectrum, ranging from non-polite expressions to grossly offensive (potentially illegal) speech. We also explore the conceptual links between offensive language and related notions such as impoliteness, hate speech and language aggression. Importantly, this chapter focuses on why understanding offensive language is, above all, a concern best addressed by linguists. To achieve this, we discuss the similarities and differences between grossly offensive and (im)polite language. We specifically focus on the pragmatic concepts such as locution, illocution and perlocution to explain how they operate at both ends of the spectrum. Finally, we address the challenges of detecting offensive language in computational approaches to tackling online hate, emphasising the importance of linguistic contributions.
Chapter 2 focuses on offensive language, positioned at the midpoint in the spectrum discussed in Chapter 1. We situate offence and the acts of causing and taking offence in the theory of (im)politeness. The chapter reviews the development of (im)politeness theory through three waves, and situates the current book within this theoretical progression. In Chapter 2, building on the previous literature, we also propose a model of offence which distinguishes between legal and moral (interactional) transgressions. To discuss the legal side, we focus on UK law, examining the challenges surrounding the concept of grossly offensive through a discussion of several high-profile cases. We also focus on moral order transgressions which occur when interactional norms are breached. The chapter also argues that taking offence, whether face-to-face or online, could be a strategic social action used to, among other things, assert boundaries or influence social norms. As we argue, online offence may function as a tool for collective action, for raising awareness and for promoting social change.
Chapter 3 explains our research methods, data collection and ethics. In this chapter, we focus on how we have addressed the typical challenges of interpreter bias in studying online offensive language. To mitigate these issues, we ground our study in naturally occurring data flagged as offensive by the targets themselves. We also introduce the reader to the world of social media content creators and discuss who the targets in our study are, and provide some demographic information about them. We discuss the use of Sketch Engine as a tool to address our primary objective of exploring the formal, lexical, semantic and discursive strategies involved in the construction of offensive language. We also discuss Wmatrix5, a corpus tool used to explore the semantic dimensions of offensive language. In addition, we explain how quantitative analysis was combined with thematic and linguistic-pragmatic approaches to examine how frequency and context shape offensive meanings. Chapter 3 further outlines our ethical considerations, including the responsible handling of data and the protection of participant anonymity throughout the research process.
In Chapter 4, we focus on the discourse of offensive language. Drawing on Foucault’s concept of discourse, we examine recurring patterns, structures and meanings that convey offensive implications in our corpus. We identify linguistic categories that are overused or underused in comparison to our reference corpus. The key features identified reveal that offensive language is often personal, opinionated and judgemental, and includes intrusive and critical expressions. In this chapter, we also explore the implications of these linguistic tendencies and their role in shaping the discourse of offence. Chapter 4 further demonstrates that the overuse of symbols, such as asterisks, serves to obscure offensive language, indicating both self-censorship and attempts to potentially normalise such language.
Chapter 5 focuses on the semantic domains of offensive language. The analysis of the semantic domains further confirms that offensive language is opinionated, intrusive and judgemental. Using the Wmatrix5 tools, we identify key areas of meaning that allow offenders to express, for example, criticism and negativity. Our analysis reveals a significant overrepresentation of terms related to, for example, dislike, foolishness and negative judgements about appearance which overall point to a tendency towards personal attacks. By focusing on these domains, we aim to address further gaps in offensive language research and predict alternative expressions of similar meanings.
Chapter 6 focuses on patterns of offensive language with a particular focus on specific words and multi-word units that characterise the discourse of offensive language. The chapter aims to enhance the classification of offensive words, refine models of offensive language, and deepen our understanding through keyword analysis. The findings reveal a wide range of explicit and taboo words, including racial slurs and derogatory terms, alongside emphatic markers, which intensify emotional statements. In our discussion of explicit language, we will critically examine the role of context in interpreting offensive language. We discuss the explicitly offensive language within the concept of ‘insults’, dividing them into two primary types: hard and soft. Although both types serve to denigrate, they differ in linguistic intensity and explicitness.
In Chapter 7 we continue discussing the patterns of offensive language by focusing on the shift from direct, explicit insults to the more indirect and implicit ones. We argue that explicitness also exists on a continuum, making it difficult to draw a clear boundary between where implicitness ends and explicitness begins. To address this, we categorise our data into more explicit and less explicit language. Less explicit language includes offensive meanings which are expressed through hinting, insinuating, implying and suggesting. As we will explain in the chapter, understanding these strategies requires considering the broader discursive context, not just the words themselves. We compare our corpus of explicit language with a less explicit one to explore the nuanced differences between the two. Our analysis reveals that themes such as gender, racial and age biases are commonly used in the less explicit corpus. As we will explain, these biases are perpetuated through seemingly polite language, rhetorical questions and comments that indirectly and subtly attack the target’s identity. Chapter 7 also discusses how offensive meanings can be inferred. Drawing on mainstream pragmatic theories such as Grice’s Theory of Implicature and Relevance Theory, we examine how implicatures arise from specific word choices and lead to various forms of offensive interpretations. We also discuss weak and strong implicatures, and how they can be used in a way that leaves room for potential deniability. Along the way, we also focus on what we call unmarked offensive implicatures – that is situations where the speaker’s intention is unclear, and the hearer assumes responsibility for interpretation. As we will explain, such implicatures could lead to idiosyncratic interpretations whereby the hearer’s personal experiences and emotional state shape their interpretation of what they see as offensive.
Finally, Chapter 8 investigates the complex relationship between online offensive language and creativity. We demonstrate how creativity, that is the ability to refashion language, extends beyond literary contexts to everyday online interactions. The chapter showcases a vast array of creative tools that are used to cause offence. The chapter as a whole attests to the fact that while creativity could be enriched by the resources available on various online platforms (e.g., memes and emojis), written language remains central to creative communication. We demonstrate how creative language tools categorised under, for example, tropes, figures of diction and figures of thought, help the offender construct remarks that target sensitive aspects such as appearance, race, gender and personality. We also argue that creative language functions on multiple levels, with its interpretation often requiring a pragmatic layer that extends beyond simple word recognition.
