Topic adversarial neural network for cross-topic cyberbullying detection.

Area of Science:

• Computational Linguistics
• Social Computing
• Cybersecurity

Background:

• Cyberbullying is a significant threat on social media, with varied linguistic expressions complicating detection.
• Existing methods often fail to generalize across different topics due to reliance on topic-specific features.

Purpose of the Study:

• To develop a novel framework for topic-invariant cyberbullying detection.
• To enhance the robustness and accuracy of automatic cyberbullying detection systems across diverse online contexts.

Main Methods:

• Proposed the Topic Adversarial Neural Network (TANN), an end-to-end framework integrating multi-level feature extraction, a topic discriminator, and a cyberbullying detector.
• Employed adversarial training to disentangle topic-related information and retain universal linguistic cues for harmful content.
• Constructed and utilized a multi-topic dataset from Chinese social media platforms (Weibo, Tieba) for evaluation.

Main Results:

• TANN demonstrated superior performance in cross-topic cyberbullying detection compared to existing methods.
• The framework significantly improved the robustness and accuracy of cyberbullying detection across different topics.
• Successfully mitigated topic interference, enabling reliable detection in dynamic online environments.

Conclusions:

• TANN offers a scalable and effective solution for topic-invariant cyberbullying detection.
• The proposed method advances the field by enabling reliable cyberbullying identification across diverse social media contexts.
• Highlights the importance of topic-invariant feature learning for robust online harm detection.