Introduction

The growing intensity of cyberbullying on social media platforms such as Twitter, Instagram, and Sina Weibo continues to stoke anxiety worldwide (Chan et al., 2021; Zhong et al., 2021). Cyberbullying not only poses grave threats to the well-being of citizens, with detrimental effects including poor self-esteem, low academic performance, and potential suicidality at the individual level (Cowie, 2013), but also impacts the social fabric. For example, it is often accompanied by negative emotions, such as anger, fear, and sadness (Xu et al., 2012a), which disrupts the positive climate of social media and public trust and beliefs towards other groups, society and government (Costello et al., 2023; Papcunová et al., 2023). Efforts to prevent cyberbullying have increased in these years. Yet it remains challenging to identify cyberbullying content within the vast array of social media texts because of its dynamic, context-sensitive and unstructured nature (Dani et al., 2017; Zhong et al., 2022). Recent progress has been made in understanding cyberbullying through the study of communication mechanisms and social behavior patterns on social media, including revealing the psychological process of different roles (Pornari and Wood, 2010; Zhu et al., 2023) and communication mechanisms on social media (Cañas et al., 2020). Building on these advancements, three research areas that necessitate finer-grained analysis are being developed: (1) the context-sensitive linguistic expression in social media, (2) the behavioral characteristics of the various roles within the cyberbullying scenario, and (3) the interplay between them. Such studies are critical to understanding the social complexity of cyberbullying and developing targeted intervention strategies.

The complexity of cyberbullying extends beyond the language to include the diverse and evolving roles involved in cyberbullying incidents (Bowler et al., 2015), which shape patterns of emotional expression. Like traditional bullying, there are three major players in cyberbullying scenarios: bullies, victims, and bystanders (Wang, 2021). Among these, bystanders (the largest group) can further be split into subcategories (e.g., González-Cabrera et al., 2019; Rathnayake et al., 2020; Y. Xu, 2021), critically shaping the evolution of cyberbullying and its impacts on victims (Polanco-Levicán and Salvo-Garrido, 2021). However, users’ roles are not static nor mutually exclusive. Besides, victims and bystanders may even respond with abusive language (Rathnayake et al., 2020). In such cases, the emotions embedded in the language of the roles become similar, highlighting the need to differentiate nuanced variations in emotions (e.g., which kind of emotion is emphasized and how it interacts with other emotions). Yet most research oversimplified the sentiment analysis (i.e., positive or negative emotions), ignoring how role-specific emotion clusters drive collective behavior (e.g., Dani et al., 2017; Choi et al., 2021; Lian et al., 2022). This obscures our focus on the emotional patterns linked to various roles in cyberbullying and, moreover, how these patterns sustain cyberbullying.

This oversight should be valued given the social effects of emotions: individuals gather relevant information from others’ emotions and make decisions (van Kleef and Côté, 2022). They tend to align their emotions with those of the group in social media (Gordijn et al., 2006; van Kleef and Côté, 2022). On social media, such a similarity in emotional responses often manifests as a role-specific pattern, helping in the prediction of emotional expression and cognitive changes in a broader group. However, it remains unclear how these roles vary with regard to emotional expression due to their complex stances and behaviors. As van Kleef & Côté (2022) suggested, behavioral implications of emotions depend on individual (e.g., identity) and situational characteristics (e.g., conflict stage), rather than emotional valence (positive or negative). This revelation exposes a critical limitation in cyberbullying research: while prior studies predominantly analyzing negative emotions (e.g., anger frequency), they glossed over the role-specific emotional profiles.

Building on this premise for viewing online roles as temporary groups, we posit that cyberbullying participants are characterized not only by role-specific linguistic features but also by similar emotions. Acknowledging the richness of involved emotions, we describe patterns of emotional expressions as co-occurrence networks of emotions and structural properties. This study first reviewed existing classifications and extracted five roles instead of only a few roles as has been done elsewhere (Jacobs et al., 2022; Rathnayake et al., 2020). Rather than focusing on the role played by individuals in social processes, as in social network analysis (Gašević et al., 2019; Choi et al., 2021), we attempted to reveal associations and patterns between multiple dynamic concepts (Shaffer et al., 2016; Rolim et al., 2019) by epistemic network analysis (ENA). We further identified how this emotional information explains the behavioral patterns of different roles based on Intergroup Emotional Theory (IET) and Emotional Empathy Theory (EASI). We believe that uncovering more granular features enriches the understanding of participants’ roles in cyberbullying and promotes more effective preventive measures. The following research questions (RQs) informed this effort:

  • RQ1: What are the distinct patterns of emotional expressions exhibited by each cyberbullying role?

  • RQ2: How do patterns of emotional expressions differ across cyberbullying roles?

Participants’ roles in cyberbullying

In the context of cyberbullying on social media platforms, users can be categorized into various participant roles according to their participation behaviors or intentions. All these roles should be valued, rather than solely focusing on the bullies and victims (Bowler et al., 2015). Bystanders play multidimensional roles in cyberbullying, ranging from constructive to aggressive intervention, which cannot be conceptualized as a unidimensional construct (Moxey and Bussey, 2019). Scholars have classified different bystanders using multiple approaches. Salmivalli (1999) argued that, in traditional bullying, bystanders span four categories: assistants of the bully, reinforcers of the bully, defenders, and outsiders. J.-M. Xu et al. (2012b) analyzed bullying on social media and pointed out three bystander groups: reporter, accuser, and other. Psychologists have labeled bystander roles based on questionnaires. For instance, Schultze-Krumbholz et al. (2018) performed latent class analysis and identified prosocial defenders, communicating outsiders, aggressive defenders, bully-victims, and assistants. González-Cabrera et al.’s (2019) categorization echoed that of Salmivalli (1999), although defenders consisted of both the victim’s defenders and supporters. Other studies have uncovered a pair of bystander roles (i.e., bystander-assistant and bystander-defender) via machine learning. It is still challenging to differentiate bullies from aggressive bystanders when detecting cyberbullying roles, as this task normally calls for additional context (Jacobs et al., 2022; Rathnayake et al., 2020).

In light of extant work, we divided cyberbullying into the following six categories, which have covered the types of roles in existing studies: bullies, victims, outsiders, the bully’s assistants (i.e., assistants), the victim’s defenders (i.e., defenders), and reporters. Outsiders consist of people who ignore or distance themselves from cyberbullying incidents. Bullies’ assistants do not initiate bullying but encourage or even exacerbate it (Salmivalli, 1999; González-Cabrera et al., 2019). Defenders are bystanders who stand up for victims or attempt to stop bullying (Jacobs et al., 2022; Rathnayake et al., 2020). Reporters refer to individuals who report bullying incidents on social media (Xu et al., 2012a) and hold a more neutral attitude towards involved parties.

Emotions in cyberbullying

Social media platform enables users to gather around common goals or identities, thereby forming temporary groups and sharing emotions (Wakefield and Wakefield, 2024). These role clusters are not maintained by formal rules or explicit goals, but by the alignment of emotions. The groups, as IET and many researchers suggested, cultivate shared emotional group norms through divergent social identities and cognitive appraisals (Smith, 1993; Smith and Mackie, 2015; Mackie and Smith, 2018), which direct further attitudes and behaviors. Specifically, the emotions of individuals playing specific cyberbullying roles will be easily affected by the intergroup, ultimately converging on a group emotional pattern for a given role (Parkinson, 2020). For example, when counter-empathic emotions (e.g., schadenfreude) are propagated between groups, they can trigger more social media engagement in groups, which may lead to social media antisocial behaviors, such as cyberbullying (Wakefield and Wakefield, 2024). In a nutshell, IET provides an irreplaceable premise for conducting role-based research, as it reveals how group identities integrate dispersed individual users into clusters of roles with stable emotional patterns.

Despite these close associations, limited studies have explored descriptive and interpretive results of discrete emotions in cyberbullying scenarios (e.g., Balakrishnan & Ng, 2022; J. Xu et al., 2012b). The commonly used categorization is Ekman’s (1992) six major types of emotions, including anger, disgust, fear, joy, sadness, and surprise and Plutchik’s (2001) wheel of emotion, which includes two additional feelings, acceptance and anticipation. Due to social media’s mechanisms, these emotions can be widely expressed and spread through various forms, including aggressive texts, images and emojis (Hettiarachchi and Ranasinghe, 2019), most of which were found to be negative, such as anger and depression (Baroncelli and Ciucci, 2014; Den Hamer and Konijn, 2016; Vranjes et al., 2017; Li and Peng, 2022).

The effects of different emotions on social media vary according to previous observations. Posts expressing intense but diverse emotions, particularly negative ones like anger, disgust, fear, and guilt (Chawla et al., 2022), tend to disseminate more rapidly and widely in both Chinese and English contexts (Stieglitz and Dang-Xuan, 2013; Zhang and Qu, 2020; Chen et al., 2023a). Such different functions can be explained using EASI. This theory posits that emotions affect cyberbullying through a dual pathway: (1) via automatic affective reactions to others’ emotional expressions; (2) via inferential processing of the meaning and implications of the emotional expresser (van Kleef and Côté, 2022). Heerdink et al. (2015) found that anger triggered more exclusion compared to neutrality and happiness. Anger for a particular event can be interpreted as a sign of injustice by others, leading to greater attributions of agency and responsibility to a third person (van Doorn et al., 2015). Furthermore, studies argue that the impact of emotions on users’ information sharing may depend on their level of arousal, rather than valence alone (Song et al., 2016; Pivecka et al., 2022). Arousal provides more information, such as urgency or importance of an incident, influencing observers’ interpretation and thus their behaviors (Storbeck and Clore, 2008; Berger, 2011; van Kleef and Côté, 2022). For example, high arousal may increase accessible judgments, including stereotyped judgments of individuals seen by raters as members of an out-group (Storbeck and Clore, 2008).

However, user posts may present multiple discrete emotions. It is unclear how the coordination of complex emotions affects role behavior. More importantly, even the same emotions work differently in different posts. For example, Jang et al. (2022) identified two groups of Twitter users who expressed negative emotions towards vaccination: those who used it to discourage vaccination, and those who promoted vaccination while criticizing the public health response. Clearly, users’ emotional patterns may vary depending on their intention and stance, or role in an incident. Quantifying the sentiment of cyberbullying posts without differentiating between participant roles can make it challenging to draw meaningful conclusions.

Although emotional patterns among roles in cyberbullying are not explicitly outlined, initial observations suggest distinct trends. Offensive roles, like bullies and their assistants, often express anger, fear, and sadness through accusatory and abusive posts, sometimes disguised with positive emotions (Choi et al., 2021; Lu et al., 2022; Zhong et al., 2022). Victims, however, may not consistently express emotions across different coping strategies, such as seeking support, defending themselves, or seeking revenge (Wright, 2016; Balakrishnan, 2018). Defenders and reporters are positive participants in a role aiming to stop cyberbullying. Defenders actively assist and comfort victims, using positive language to encourage them (Nickerson et al., 2008; Alim and Khalid, 2019). Reporters, on the other hand, may express sadness, sympathy, neutral or other negative emotions to drive the discussion or question the responsible parties (Gu et al., 2021). Outsiders tend to respond more passively or apathetically, reflecting a more objective stance (Jia et al., 2022).

Methods

We applied a four-phase method to explore the emotional characteristics of cyberbullying roles on Chinese social media, as shown in Fig. 1.

Fig. 1
figure 1

Research framework.

Full size image

Data collection and preparation

Our experiment utilized a corpus of three representative cyberbullying incidents from China’s largest platform for open discussions, Sina Weibo. With mechanisms similar to Twitter, it provides clear hashtags related to an incident, which is ideal for gathering relatively comprehensive data on cyberbullying (Li, 2020). The chosen cases coincided with prior criteria (Zhong et al., 2022), concerned three spheres of daily life: society, sports, and entertainment. The societal case involved an ambiguous sexual assault allegation, surrounded by rumors about women’s safety. The sports case involved gender-biased remarks towards a sports celebrity, sparking public dialog on gender issues. Lastly, the entertainment case was a hate attack triggered by an entertainment star’s depression diagnosis. These incidents have been among the most widely discussed on Sina Weibo over the past five years and have ignited considerable public concern (with over 140 million reads and over 30,000 discussions). Despite contextual differences, they represented typical cyberbullying topics that can be easily found across cultures. They then exhibited ambiguity that drew diverse audiences into discussions, improving the complexity and role diversity of cyberbullying. By analyzing roles across these cases, this study could identify common emotional patterns in cyberbullying, undetectable in isolated case studies, with limited data.

A Python-based crawler was developed to jointly download posts and comments about each incident. Microblog text, usually containing multimedia information, requires special handling for data storage and annotation. We thus established a cleaning pipeline including: (1) removing duplicates, whitespaces, hyperlinks, unknown characters, user mentions, and HTML tags; (2) converting traditional Chinese into simplified Chinese; and (3) retaining emojis with matching words. Eventually, we assembled a raw corpus comprising 11,601 posts.

The data used in this study were collected exclusively from publicly accessible posts and comments on Sina Weibo, in compliance with the platform’s Terms of Service. To protect user privacy, all identifiable information (e.g., usernames, profile links, geolocation tags) was removed prior to analysis, and aggregated results were reported to prevent re-identification. No content from private accounts or restricted communities was included. Data collection protocols adhered to institutional guidelines for ethical use of public social media data.

Cyberbullying roles and human coding

As mentioned, we named six cyberbullying roles, all of which represented illocutionary acts based on the posts’ context. Victims’ emotional patterns were not analyzed. Victims are typically defined as individuals “who are the target of repeated harassment” (Salmivalli, 1999) or who experience being harassed. We viewed these persons as the heart of an incident (i.e., by being the people who were initially bullied) rather than as persons who others verbally attacked during an incident. Victims’ posts would also have been difficult to process because few victims regularly speak out on social media.

Rather than coding at the user-level, our coding was conducted at the post-level, i.e., the coding of a complete comment, excluding the commenter’s account details for privacy reasons. That is, in our study, cyberbullying roles were categorized based on their behavioral intent reflected in individual posts rather than being statically assigned to user accounts. Based on our coding scheme (listed in Table 1), five independent coders with a uniform understanding of the three cyberbullying incidents were trained to determine which role each post reflected. Specifically, we implemented a three-stage strategy for this manual coding task. In the first stage, we conducted a two-round consistency test, which included coder training and intercoder agreement discussions. We specifically used Holsti’s coefficient to evaluate inter-rater consistency (Holsti, 1969), which accounts for both agreement and disagreement across multiple coders and categories. It is considered to be useful when coding requires holistic judgments of context, such as in discourse analysis or frame interpretation (Nili et al., 2020). The coefficient in the second round was 0.93, indicating high consistency among annotators. In the second stage, each coder was assigned to annotate 2320 posts independently over a one-month period. The coding proceeded sequentially by case, with posts from each case randomized before being presented to the coders. In the final stage, we conducted circular member-checking, where each coder was assigned to cross-check the annotations of another coder in a closed loop. This approach not only reduced individual workload but also mitigated potential biases, enhancing the overall quality of the annotations. Our final dataset contained 11,601 labeled instances. Table 2 displays the basic statistics for these data.

Table 1 Coding scheme for roles on Sina Weibo.
Full size table
Table 2 Statistics for Sina Weibo cyberbullying dataset.
Full size table

Fine-tuning ERNIE 3.0 for emotion detection

To handle social media posts with multiple emotions, we treated emotion detection as a multi-label classification task for short text. Challenges included the exponential explosion of label combinations and the computational costs associated with constructing brief-text classification models (Gibaja and Ventura, 2015). Few studies have extracted diverse emotions from Chinese social media in this manner. However, recent advances in text-based emotion detection, particularly with transformer-based models (Acheampong et al., 2021), have shown promising results with improved efficiency and reduced dependency on labeled data. We therefore fine-tuned a transformer-based model to enhance fine-grained emotion detection.

This goal was realized through three steps: (1) selecting a transformer-based model; (2) feeding a corpus of social media posts into this model for fine-grained emotion detection; and (3) evaluating emotion detection from social media. The ERNIE 3.0 pretrained model (Zhang et al., 2019; Sun et al., 2021) is founded on auto-encoding transformers. It was trained using a 4TB corpus consisting of plain text and a large-scale knowledge graph. Compared with BERT and GPT, ERNIE 3.0 has demonstrated superior performance on Chinese NLP tasks. We fine-tuned ERNIE 3.0 on the public corpus of Wang et al. (2014) to adapt its parameters for social media emotion detection. The fine-tuned ERNIE 3.0 was then employed to classify multiple labels in our sample using a simple discriminant function. A pragmatic assumption of this function is that an emotional label will be assigned to a social media post when its predicted probability is greater than the threshold of 5% (Storey and Tibshirani, 2003; Read et al., 2011; Liu and Chen, 2015; Alotaibi and Flach, 2021). To ensure consistent performance of the fine-tuned ERNIE 3.0 in our research context, we implemented a domain adaptation framework inspired by Ben-David et al.‘s (2010) theoretical work on dataset shift mitigation. We constructed an independent test set rigorously partitioned from our dataset (stratified 20% split, n = 2320). Consequently, evaluation on the test set revealed encouraging results with an average precision of 77.7%, a recall rate of 77.2%, and an F1 score of 77.3%. We fully acknowledge that an over 20% misclassification rate might introduce certain biases in the analysis of role-specific emotional patterns. To address this, we performed bootstrap resampling with 1000 iterations on our dataset to construct 95% confidence intervals. Results confirmed that the significance levels (p < 0.05) of emotional differences between roles remained stable despite misclassification errors (Appendix has more details). This indicates that the misclassification did not fundamentally alter the statistical significance of our findings, thus not leading to significant discrepancies in the conclusions and ensuring the robustness of the results.

Data visualization and analysis

After fine-grained emotion detection, we statistically analyzed the data and applied ENA to address our RQs. For RQ1, we first summarized the counts of emotional expressions for each role and plotted an “affective flower” chart. Specifically, to mitigate potential bias from highly active users, we calculated the proportion of each emotional expression type relative to each role’s total post count, thereby normalizing frequency measurements and reducing overrepresentation risks. Then, we performed a pairwise multiple comparison test, Conover’s test, to uncover differences in emotional expressions across roles. Here, basic emotions and cyberbullying roles were regarded as dependent variables and independent variables, respectively. This is one of the most powerful procedures for post hoc evaluation (Conover and Iman, 1979) and has been implemented in a Python package, scikit-posthocs (Terpilowski, 2019).

For RQ2, ENA was exploited to detect fine-grained differences that existed in emotional patterns among roles. ENA is capable of analyzing discourse data in groups and revealing associations between multiple dynamic concepts (Rolim et al., 2019; Shaffer et al., 2016). This capability contributes to its popularity for extracting cognitive patterns from social media data. Misiejuk et al. (2021) theorized that integrating sentiment analysis in ENA allowed for a clearer understanding of how groups saw the components of a stimulus issue by separating the elements by sentiment. We entered our data into the ENA web tool. First, we chose six fundamental emotions as codes, with each cyberbullying role serving as a unit of analysis. The stanza was then set to 1, given our interest in cumulative connections among these emotions. Nodes within our epistemic network represented basic emotions, and link thickness indicated co-occurrence frequency; the thicker the link, the more often the emotions co-occurred.

Results

Distributions and differences in emotions among cyberbullying roles

To discern the distributional characteristics of cyberbullying roles among the six emotions, emotion analysis was performed on posts typifying each role to analyze the associated feeling. Figure 2 depicts the results. The value in the center of an affective flower represents the emotion evenness index (EEI; Huang et al., 2020). Basic emotions were evenly distributed within a specific role. Each flower’s oval petal represents the number of posts in which a particular emotion appeared. Expressions of the chosen emotions varied across roles. Unsurprisingly, Angry was most pronounced among bullies, with a similar pattern among assistants and reporters. Surprised, Neutral, and Fearful were most common among outsiders. Happy was prevalent for defenders, followed by Angry and Neutral. Sad ranked the lowest across all roles except for reporters.

Fig. 2
figure 2

Distributions of emotion among five roles.

Full size image

As shown in Fig. 3, Conover’s test was carried out to ascertain potential significant differences between roles tied to these emotions. For Fearful, Happy, and Surprised, all roles exhibited a significant difference (p < 0.05) except for bullies and assistants. Significant differences between roles were largely apparent for Angry, Fearful, Happy, and Surprised; fewer pairs of significant differences accompanied Neutral. Regarding Angry, no significant difference was observed in bully–assistant and reporter–assistant pairs. Significant differences in Sad were observed for most pairs, except for bully–assistant, outsider–defender, and outsider–reporter. Variation was also evident between outsiders’ neutral emotions and those of other roles.

Fig. 3
figure 3

Comparisons of emotional patterns among roles.

Full size image

Epistemic networks of emotions for cyberbullying roles

Overall characteristics of epistemic network

Figure 4 shows the overall network for the six emotions based on five roles. Along the x-axis, the space is distinguished by Angry on the left and Happy on the right. The y-axis is defined by Surprised at the bottom and Angry at the top. Based on these distributions of emotions and roles, the y-axis may represent how strongly different emotions contribute to cyberbullying behaviors, distinguishing more active roles from passive ones. For example, the roles situated above the x-axis are more aggressive (i.e., showing Angry), being the major contributors of cyberbullying, while outsiders who were not involved in cyberbullying show the lowest y-values. It seems that the x-axis is related to emotional valence (Berger and Milkman, 2012), where most negative emotions are on the left and positive emotions on the right, except for Sad. Dots denote emotions, and lines denote connections between these feelings. Larger dots and thicker lines indicated more prominent emotions and stronger connections, respectively. In the overall epistemic network, connections between Angry and other discrete emotions (particularly Surprised, Neutral, and Fearful) were stronger compared to other emotion pairs (e.g., Sad-Neutral), evident from the concentration of lines on the left side of Fig. 4.

Fig. 4
figure 4

Mean network graphs of emotions for five cyberbullying roles.

Full size image

Figure 5 visualizes the centroids of different roles in the ENA space. These centroids’ average values are displayed as squares, and the 95% confidence interval for each dimension is shown as a rectangle. Each centroid considers the weights of connections between emotions and can be depicted as a matching plotted point (Bressler et al., 2019). Obviously, the co-occurrence pattern of conflicting emotions among defenders (red) resulted in a greater occupancy in the ENA space with the centroid to the left of the x-axis. Conversely, the emotions of outsiders (purple) appear to be concentrated in a smaller area. Moreover, the centroids of bullies (yellow) and assistants (green) nearly overlap, with similar confidence interval sizes, suggesting a high degree of consistency in their emotional expressions.

Fig. 5
figure 5

Mean of plotted points for five roles.

Full size image

Figure 6a–e present each role’s epistemic network. The lines link different points (emotions), such that the thicker the connecting line, the greater the co-occurrence frequency of the associated emotions. Bullies, assistants, and reporters similarly expressed emotions stronger associations between negative emotions, whereas reporters were less associated with Happy than the other two roles. Specifically, defenders primarily displayed a conflicting emotion co-occurrence pattern. Outsiders tended to express themselves neutrally.

Fig. 6: Networks of co-occurring emotions across different roles.
figure 6

Each panel shows the epistemic networks of a specific role: a Bullies, b Outsiders, c Defenders, d Assistants, and e Reporters.

Full size image

Differences in role categories

When examining networks, it is useful to compare a pair of network models by subtracting their connection weights from each other to create a difference network graph (i.e., a “subtracted network”). Our graph provided additional insight into the distinctions among cyberbullying roles, as pictured in Fig. 7.

Fig. 7: Subtracted networks of emotions for different roles.
figure 7

Panels show ten pairs of substracted networks, including: a Bully-Outsider, b Bully- Defender, c Bully-Assistant, d Bully-Reporter, e Outsider-Defender, f Outsider-Assistant, g Outsider-Reporter, h Defender-Assistant, i Defender-Reporter, and j Assistant-Reporter.

Full size image

The subtracted networks illustrate emotional differences and co-occurrences among roles, shedding further light on parties’ behavioral motivations around cyberbullying. For brevity, we focus on three key pairs that showed the most significant differences in all discrete emotions except Neutral: bully–assistant, defender–assistant, and outsider–reporter. Figure 7c shows that the subtracted network of bullies and assistants had thin edges, implying a slight difference between their networks. The bullies’ emotional network contained stronger connections between Angry and Sad, while the assistants’ network had stronger links between Angry and Fearful, Angry and Neutral, and Fearful and Neutral. Figure 7h reveals that the defenders’ network mainly occupied the right side of the axis, with stark connections between Angry and Happy and Neutral and Happy. Defenders’ emotions were generally positive, while assistants frequently appeared Angry and Fearful or Surprised. Figure 7g indicates that outsiders’ emotions were easily distinguishable from those of reporters. Angry and Surprised, and Angry and Fearful appeared more frequently in the defenders’ network. Neutral and Surprised, Fearful and Neutral, and Fearful and Surprised were more common among outsiders.

Discussion

The picture of emotions in cyberbullying: more than anger and valence

The emotional expressions of different roles are complex and multifaceted. Angry, Fearful, Happy, and Surprised varied significantly among nearly all roles, while Sad was comparatively less prevalent and less likely to vary significantly among roles. These findings align with the popular view of Angry, Fearful and Sad on social media (Xu et al., 2012a; Baroncelli and Ciucci, 2014; Li et al., 2020; Chawla et al., 2022; Zhong et al., 2022).

Our overall epistemic network further indicated that emotional valence may be an unreliable indicator of cyberbullying, which is consistent with research by van Kleef & Côté (2022) that the positive or negative impact of emotions depends on individual and situational characteristics, not just their valence. Findings suggested that Angry, which exhibits high levels of arousal, contributed prominently to cyberbullying. Conversely, low arousal emotions such as Sad and Surprised had a weaker association with cyberbullying. Thus, we presume that the arousal level of emotions may also be linked to core and non-core roles in cyberbullying. Emotions with higher valence and arousal were found to last longer on microblog (Sener et al., 2023). Anger is considered a dominant high-arousal negative emotion, leading to more engagement in harassing, bullying or hurting perceived threats, whereas fear may result in passive, harmful responses, such as ignoring and avoiding these threats (Clobert et al., 2022). As for the unexpected result for Sad, it could be attributed to posts filled with sadness tend to be less shared and disseminated on social media (Chawla et al., 2022). Therefore, future research should broaden its focus to include other high-arousal emotions like Fearful and Happy, and explore their correlations.

Three types of emotional patterns: bridging IET and EASI

As reviewed above, IET posits that shared social identities (e.g., ‘bully’ as a transient identity) bring emotion convergence on social media (Mackie and Smith, 2018). This convergence, we argue, implies that different discrete emotions may manifest distinct patterns of co-occurrence with different roles. The phenomenon of group emotional convergence then tends to strengthen internal cohesion, allowing group members to develop a consensual interpretation of future events (Smith and Mackie, 2015). This essentially involves the mechanism proposed by EASI—whether the users will have automatic affective processing or deliberate inferential analysis. Notably, the type of emotion may affect the depth of processing (Tiedens and Linton, 2001). Studies indicated emotions associated with uncertainty, such as sadness, produced more thorough processing, while anger emotions with clear attributional directions may result in heuristic processing and stereotype use (Tiedens and Linton, 2001; Small and Lerner, 2008; Tugade et al., 2014). Together, we believe these two theories provide a useful lens for decoding behavioral patterns across different cyberbullying roles.

In the networks of emotion co-occurrence among different roles, three types of patterns were identified, including an anger-dominated negative pattern, a happiness-anger conflicting pattern, and a surprise-fear moderate pattern:

  1. (1)

    Anger-dominated negative pattern was observed in bullies, assistants and reporters, with Angry being the most prevalent, appearing alongside other negative emotions, in line with previous studies (Mackie et al., 2004, p. 200; Baroncelli and Ciucci, 2014). Reporters, as interveners, may perceive crisis incidents on social media as threats to fairness (Zhang et al., 2018), forming group norms emphasizing condemnation of injustice. They might vent negative emotions, especially anger marked by clear attribution and high arousal level. They enable reporters to rapidly assign responsibility and advocate accountability (Small and Lerner, 2008), as reflected in their concise and punitive comments. In contrast, threat-induced fear and anxiety, characterized by low certainty and minimal emotional polarization (Renström et al., 2023), conflicted with their action goals.

  2. (2)

    Happiness-anger conflicting pattern was found in defenders. This finding correlates with the two defender behavioral patterns observed in the corpus. From an IET perspective, defenders may simultaneously contain two subgroups with different norms: one required to stand up for the victim and the other to confront bullies. Specifically, the former may adopt a strategy of solidarity with victims, conveying happiness or empathy for emotional support, while the latter may express anger to advocate for justice and counteract cyberbullying (Vranjes et al., 2017; Schultze-Krumbholz et al., 2018). This dual identity could cause conflicting emotions, reflected in the largest confidence intervals in the ENA space. Furthermore, defenders frequently employed sarcasm over direct aggression, reflecting Chinese social media’s prevalent implicit cyberbullying strategies of mockery and ridicule (Zhong et al., 2022).

  3. (3)

    Surprise-fear moderate pattern, which presented more Surprised, Neutral and Fearful, with these emotions showing co-occurrences with Happy and Angry, respectively. This pattern was identified in outsiders who were conceptualized as neutral observers in this study. They regularly expressed surprise at aggressive comments due to not expecting such behavior. Mameli et al. (2022) came to similar conclusions. According to IET (Mackie and Smith, 2018), their emotional expressions were more balanced, rather than convergent, as they were unaffiliated with either group involved in an incident. These detached identities enabled outsiders to keep themselves from being emotionally drawn into a particular group. Instead, they expressed a certain degree of rationality (as indicated by neutral emotions).

Fine-grained variations among cyberbullying roles

We compared each pair of roles and highlighted three pairs with similar social effects (i.e., bully-assistant, defender-assistant, and outsider-reporter). The findings indicate that analyzing emotional co-occurrences can reveal the positive or negative social effects of emotions in distinct roles, as well as their behaviors and intentions, aiding in role identification. This differentiation helps to distinguish between similar roles.

Key findings revealed significant emotional overlap between bullies and assistants, as evidenced by centroid proximity and confidence interval convergence. Drawing on IET and EASI (Mackie and Smith, 2018; van Kleef and Côté, 2022), this alignment may arise from shared group identification: assistants, by supporting bullies, adopted similar group norms and maintained a unified stance toward the same goal. This also explains why interventions targeting only bullies or haters often fail, as the emotional pattern persists through assistants’ unconscious compliance. Nevertheless, their different co-occurrence between Angry and other negative emotions facilitated their differentiation. This result may be due to their different aggressive behaviors. Sad in bullies’ epistemic network may stem from public reprimand for aggressive behavior by one party to the incident (Pornari and Wood, 2010) or from realizing someone had been hurt (Mameli et al., 2022). In contrast, assistants’ anger often co-occurred with Fearful, Surprised, and Neutral, as they usually conceal themselves within groups, projecting objectivity and impartiality but in fact share alarming comments to encourage aggression and cyberbullying (Snakenborg et al., 2011).

Being supporters of the victim and cyberbullying, respectively, comparing the epistemic networks of defenders and assistants offered valuable insights into two distinct approaches to cyberbullying intervention. The assistant displayed an aggressive style, which increased conflict by negative emotions, while defenders showed more pro-social behaviors by comforting and explaining to the victims. Previous studies revealed that expressing more pleasant emotions can promote positive social effects, such as pro-social bystander behaviors (Miles and Upenieks, 2018; García-Vázquez et al., 2020; van Kleef and Côté, 2022). Therefore, their epistemic network mainly occupied the right side of the axis with clear connections (i.e., Happy-Angry and Happy-Neutral), suggesting predominantly positive emotions. On the contrary, unpleasant feelings about cyberbullying expressed by assistants compel people to support victims (Costa Ferreira et al., 2022). This transmission of anger may boost assistants’ level of influence and promote cyberbullying (Barsade, 2002; Sarmiento et al., 2019).

As for the outsider-reporter comparison, although they both engaged in the cyberbullying events as observers, their emotional co-occurrences varied. For outsiders, Neutral, Surprised, and Happy have a strong co-occurrence between them, while reporters’ emotional responses were predominantly negative, consistent with J.-M. Xu et al.’s (2012b) study. Except for the low perceived threat, some users may be afraid to post negative comments about others and become the target of bullies’ revenge. These circumstances can impede outsiders from intervening in cyberbullying (Wang, 2021; Costa Ferreira et al., 2022). Reporters were apt to express anger in order to motivate others to exert more cognitive energy attending to the unjust news (Stevens et al., 2021). Such expressions also have the potential to instigate change or extract concessions in conflict (van Kleef and Côté, 2022).

From patterns to interventions: suggestions for cyberbullying prevention

These emotion-based patterns not only deepen our understanding of interpersonal dynamics in cyberbullying but also offer some practical suggestions for prevention. The high involvement of anger-based negative emotions in most roles has again highlighted the importance of anger (Rathje et al., 2021). However, focusing solely on anger may overlook posts from defenders and reporters, whose roles could differentially influence the trajectory of cyberbullying. In addition, the discussion on the similar emotional patterns exhibited by assistants and bullies indicated that effective interventions must weaken the shared stance that binds them into a cohesive unit. As noted earlier, prior cyberbullying research has predominantly focused on anger detection in isolation. Consequently, automated detection systems often demonstrate suboptimal performance owing to their inability to incorporate crucial contextual cues. The observation of co-occurrence patterns between anger and other high-arousal emotions suggests that emotion pairs—rather than isolated emotions—could serve as potential features for identifying cyberbullying content.

Conclusion and implications

This paper has presented a novel exploration of emotional patterns among five roles in cyberbullying scenarios. We leveraged the ERNIE 3.0 model for emotion detection and ENA for data visualization. Our analysis of 11,601 comments about three cyberbullying incidents on Chinese social media revealed three emotional patterns among five cyberbullying roles: bullies, assistants and reporters presented an Anger dominated negative pattern, namely Angry co-occurred with Surprised and Fearful; defenders showed a Happiness-Anger conflicting pattern and outsiders displayed a Surprise-Fear moderate pattern with balanced distribution of different emotions in their networks. Furthermore, although the overall epistemic network featured more robust connections to negative emotions for all groups, participant roles in cyberbullying may not only depend on valence but also be related to arousal level. Discussions were presented on the basis of IET and EASI, which collectively explained why roles developed distinct emotional patterns (IET) and how these differences correlated with role-specific behaviors (EASI). Ensuing fine-grained analysis of the subtracted networks for three role pairs (i.e., bullies-assistants, defenders-assistants, and outsiders-reporters) showed variations in these pairs’ emotional co-occurrence, facilitating further understanding of their perceptions, inferences, and intentions related to cyberbullying incidents.

This study offers valuable insights into the diverse roles in cyberbullying, contextualizing interpersonal interaction through emotional patterns (e.g., how aggressors may respond to others or how bystanders evaluate ongoing incidents). Our application of two theories to role identities in cyberbullying can also transfer to other online contexts, such as emotionally polarized political discussion and crisis-driven public debates. Moreover, the findings also highlight the role of emotions in cyberbullying prevention and governance. For instance, interventions should not solely target bullies; assistants, who often mimic and thereby extend bullies’ negative emotional influence, require attention as well. A psychosocial approach, considering affective factors, may inform targeted preventative measures. Lastly, further studies can focus on aspects of group cohesion and collective cognitive evaluation in cyberbullying and establish associations with intense co-occurring emotions rather than discrete emotions.

Limitations

The current study has several limitations. First, our dataset is limited in its ability to capture the full spectrum of cyberbullying variations, including cross-cultural and cross-platform differences, where larger and more diverse samples are often required. This suggests the need for enhancing AI-assisted methods to balance data size and reliability. Implementing hybrid human-AI validation frameworks could strengthen emotion classification reliability (Chen et al., 2023b). Besides, it must be acknowledged that the combination of human and machine methods does not preclude the possibility of a single post with multiple intentions, and that misclassification of the model would affect our conclusions to some extent. Third, our application of IET and EASI remains at a theoretical level. Empirical validation, such as time-series analysis or longitudinal tracking of individuals, is needed to examine specific causal mechanisms. Nonetheless, the findings of this exploratory study offer a foundational framework for tackling real-world intervention challenges.