Emotion-Adaptive Training for Cyberbullying Detection: Bridging the Gap Between Harassment and Defamation in Celebrity Cases

name-calling, threats, or attacks—there’s a growing concern about more subtle forms of bullying, like defamation and spreading rumors, especially when it involves celebrities. Traditional cyberbullying detection systems, which tend to focus on explicit harassment, struggle to capture these indirect forms. This gap leaves a critical issue unaddressed, particularly as public figures are often targets of such complex, insidious online abuse.

In this study, researchers sought to fill this gap by creating a specialized dataset, HDCyberbullying, that focuses on cyberbullying aimed at celebrities, capturing both harassment and defamation. To improve detection, they explored transformer-based models like RoBERTa, BERT, and XLNet, but faced challenges when attempting to distinguish between the two types of bullying. The solution? The Emotion-Adaptive Training (EAT) framework, which enhances cyberbullying detection by leveraging emotion detection data. The results were promising, showing a 20% improvement in model performance.

Understanding the Core Concepts:

• Cyberbullying: Cyberbullying involves the use of digital platforms to harass or mistreat others. It can be direct, such as verbal threats or insults, or indirect, like defamation, spreading rumors, or creating false narratives. Detecting these forms, particularly defamation, is critical as they can have lasting psychological effects on the victim, especially when directed at celebrities.

• Emotion-Detection Domain: This refers to the identification of emotions, such as anger, sadness, and disgust, within text. Emotion detection is valuable in identifying the underlying tone of online content, which is often indicative of harmful behaviors like harassment or defamation. By transferring knowledge from emotion detection to cyberbullying detection, models can better identify the more subtle forms of online abuse.

• Low-Resource Settings: Many applications in the realm of cyberbullying detection face the challenge of limited labeled data, making it difficult to train robust machine learning models. The EAT framework addresses this by utilizing emotion detection data, which is often more abundant, to help overcome the scarcity of cyberbullying-specific datasets.

The Challenge: A Lack of Suitable Datasets and Low-Resource Settings

One of the key challenges in cyberbullying detection is the lack of suitable datasets, particularly those that capture the full spectrum of bullying behaviors. Most existing datasets focus heavily on direct forms of harassment, such as name-calling and threats, while neglecting more complex forms like defamation. For celebrity cases, where public figures are often the subject of rumors and false information, current datasets fall short.

Furthermore, many of these datasets are not designed to tackle the issue of low-resource environments, where labeled data is sparse. This made it difficult for previous research to build effective models for identifying the subtle nuances of cyberbullying, especially in the case of celebrities.

To address this gap, the researchers created the HDCyberbullying dataset, which includes labeled instances of both harassment and defamation targeting real-life celebrities. However, the dataset still faced challenges such as imbalances in text length and class distribution, with defamation examples being less frequent than harassment.

Emotion-Adaptive Training (EAT): A Novel Approach to Transfer Knowledge

To tackle the complexities of classifying both harassment and defamation, the researchers introduced the Emotion-Adaptive Training (EAT) framework. The core idea behind EAT is to use emotion detection datasets—often more diverse and abundant—to train models, then transfer that knowledge to the cyberbullying domain.

Here’s how it works:

• Training on Emotion Data: First, models are trained on emotion detection datasets, which are rich in examples of text that express various emotions, such as anger, sadness, and disgust.

• Transferring Knowledge: After training on emotion data, the model’s learned representations are adapted to the cyberbullying detection task. This transfer enhances the model's ability to differentiate between direct harassment and indirect defamation.

• Adaptive Learning: The EAT framework uses an adaptive approach to gradually adjust the importance of emotion-related features. This ensures a smoother transition between the emotion detection and cyberbullying detection tasks, reducing the capacity gap between models.

Experimental Results: EAT’s Effectiveness Across Multiple Models

The EAT framework was tested across several transformer-based models, including RoBERTa, BERT, and XLNet. The results were clear: EAT significantly improved performance, particularly in multi-class cyberbullying detection. The framework led to an average 20% improvement in precision, recall, and F1 scores, making it particularly effective in identifying the difference between harassment (direct bullying) and defamation (indirect bullying).

The enhanced models were better equipped to detect subtle forms of cyberbullying, such as defamation, which has always been a challenge for traditional models. This is especially important in the context of celebrity cyberbullying, where defamation often plays a central role.

Theoretical Insights: How EAT Works

The success of the EAT framework can be explained through transfer learning and domain adaptation. The idea is that the emotion detection domain shares significant features with the cyberbullying detection domain, even though the datasets differ. By training on emotion data first, the model can adapt its learned features to better capture the nuances of cyberbullying, especially the indirect forms like defamation.

Moreover, the framework uses an adaptive process that gradually shifts focus toward cyberbullying detection as the model learns, ensuring that it doesn't overlook the subtleties of the different types of bullying.

Conclusion: Bridging the Gap in Cyberbullying Detection

This study represents a major leap in the detection of celebrity cyberbullying. By introducing the HDCyberbullying dataset and the Emotion-Adaptive Training (EAT) framework, the researchers have made significant strides in identifying both harassment and defamation, which have typically been treated separately. The results show that combining emotion-adaptive training with transformer-based models can substantially improve detection capabilities, especially in low-resource environments.

The approach has far-reaching implications for social media platforms, helping them better identify and mitigate both direct and indirect forms of abuse. With the rise of cyberbullying targeting celebrities, this method could be a crucial tool in improving online safety.

Given the rise in cyberbullying, particularly among high-profile figures, how do you think the emotion-adaptive approach can influence the future of social media moderation? Could it help platforms identify more subtle forms of online abuse, ensuring better protection for users?